Inspired by a spate of recent announcements, I want to explore a thought experiment. If the price of robots keeps dropping, and their capability keeps going up, how would the world change?

The widespread success of coding models has made software easier to create than ever. You can now, with a modest investment, create your dream website or app — without being held back by years of software engineering experience. As we move towards a world where software is “free” — where frontier models like Fable 5 can one-shot a whole app — we must start wondering about what happens if we apply the same logic to robotics and, therefore, labor.

The fundamental premise of robotics is that we can automate large swathes of human manual labor and replace it with machine labor. In software engineering, abundance means more and better websites, easier data analysis, and more apps for niche applications. For robotics, universal automation — “free,” or really just radically discounted labor — means the “long tail” of jobs that just are not even worth automating right now.

What does this work look like?

Subscribe now

The Work Not Done

To understand the sheer amount of work not being done, all you need to do is to look around.

Maintenance and cleaning are constantly postponed until the last possible minute. With infinite labor, we can do actually preventative maintenance: re-grout and re-caulk tiles, repaint road lines every month so they always look brand new. Gutters, roofs, and weatherstripping can be re-done before structural issues form.

Litter would vanish. Public spaces could be maintained, by robots, at the level of a boutique hotel. The world can be made clean and beautiful, with gardens and public art and clean streets everywhere.

And it goes past this - with free labor, homes can be de-cluttered and automatically organized, their contents entered into a contextual database, instantly queryable and searchable, accessible to AI agents to make sure that, for example, the ingredients for dinner are ready and prepped the second you get home.

Things can be made to last longer. We have a culture which consumes endless, disposable goods in part because of cost of labor: a modern industrial production line can mass produce new goods, but always the same goods. What’s worse, repairing old ones requires expensive human hands — so we just don’t do that. With “free” robotic labor available at scale, we might repair old shoes or upgrade old phones instead of throwing them out, reupholster old furniture, and patch worn clothes instead of consigning them to a landfill.

Food and goods can be made incredibly customizable. Every piece of clothing custom fit; every meal freshly-prepared from fresh ingredients.

Even care work can be made easier. Even if the human elements of teaching, childcare, and eldercare aren’t replaced, free labor makes the burdens of such jobs lighter — robots could reset classrooms, prepare activities, and help carry heavy weights.

Jevons Paradox

One must then wonder if, in such a world, there will be less jobs.

Fundamentally, though, we must see that the demand for things is essentially infinite. With nigh-infinite labor, there is a massive amount of work that can be done, and therefore demand for labor will likely skyrocket.

This has been studied before, as Jevons Paradox, and it has occurred many times in human history. As travel got cheaper thanks to the automobile, people traveled more; as lighting got cheaper, thanks to the electric lightbulb, human societies generated more and more and more light.

The classic example of Jevons paradox in automation is the story of the bank teller and the ATM. Originally thought to replace bank tellers, the invention of the ATM actually saw the number of bank tellers increase until the 2010s. Human bank tellers could focus on more high-valued work, leading to the opening of new branches.

Likewise, we might think of this “free labor” as replacing not workers but enabling them — there will remain many decisions to be made, about what should be built and what work should be done. Doing the right thing — having taste — will become more important than rote labor.

Why This World Might Be Coming Soon

Good, useful robots are getting much, much cheaper, something I have written about before here. The fundamental barrier is no longer cost, but intelligence: can we make these cheap robots reliable enough, at a wide range of tasks, so that we can make this explosion of labor actually useful to people.

The other thing worth considering is that this both builds upon, and enables, an expansion of agentic AI. Endless, cheap, robotic real-world labor will be the hands and eyes and mouths of cloud-based AI, accumulating data and creating the context for future decision making.

For all of this to happen, we just need current trends to continue to hold. Hardware needs to get cheaper, robots need to become more plentiful, and AI needs to continue to get more powerful and more general as it can ingest more data.

From my point of view, this is practically guaranteed.

Subscribe now

Share

Leave a comment

Read More

When Robots Create Jobs by Nils of Auki

A couple startups have announced really exciting general-purpose robots in the sub-$10,000 range. These are finally, finally starting to look like something that could be made very capable.

This is exciting because it means affordable home robots, easily in the same range as common home appliances like a refrigerator or a dishwasher, are clearly only a few years from the market.

Let’s go over a couple of the latest announcements.

Subscribe now

Nori L2

Size: 100cm or 130cm tall for Grande (3-4 feet)

Price: $1,288 or $1,322 for the Grande model

Compute onboard: no (thin client)

Nori Robotics, founded by Antonio Li, is selling possibly the cheapest reasonably large bimanual robot I have ever seen. It’s tall enough to reach the top of a table or countertop (though definitely not a top shelf).

The arms look similar to the SO-100, a very common low-cost open-source arm by HuggingFace; probably nice if you want to build a leader setup for teleoperation for a few hundred extra dollars. The robot itself also has a nice screen useful for telepresence.

You can read about an earlier version of in the Nori robot paper, which probably owes some inspiration to the XLeRobot which I have covered before.

You can see the announcement video here.

BracketBot 2.5

Size: ~5’ 8”

Price: < $3000

Compute onboard: possibly a Jetson Nano

Another wheeled robot manipulator, this one with a distinctive tall, thin shape, and hoverboard-style drive wheels (with stabilizers running out in front). The creators want to ship “a billion robots” and with a price like this it seems totally reasonable.

There’s a blog post on a similar project here, written by Clayton Haight. Bracketbot, though, is founded by Brian Machado.

Learn more in the X post.

Weave Isaac 1

Size: probably roughly ~5’ 6”

Price: $8,000 or $450/month

Compute onboard: unknown; teleop-enabled

The successor to Isaac 0, the clothes-folding home robot, Isaac 1 is a fully capable home mobile manipulator which can put away laundry and tidy a room. This is the first robot promising the full (semi-) autonomous experience that you can pre-order this year, after last year’s launch of the 1X NEO for pre-orders.

This probably has the most robust and full-featured software stack; you can already see Isaac 0 robots in different locations, folding clothes. According to Weave, the older model folds 1000+ pounds of clothes every week.

Another note: unlike the other two, this one clearly comes with a charging dock, which is a great and really underrated feature for a robot like this.

You can pre-order Isaac 1 here.

Final Thoughts

These robots have a couple similarities that are letting them run at low cost.

• Wheels over legs. Wheels break less and require less parts. The tradeoffs in terms of going up and down stairs will most likely be worth it in terms of product lifespan without spending much more on software, sensors, compute, and model training.

• Remote inference. None of them seem likely to have much compute for models; they’re largely going to rely on the cloud for this, or some other local offboard compute. Isaac 1 may be the exception, but it does allow for teleop.

• End effector cameras, but simple grippers. No fancy tactile sensors or dexterous hands - just robust hardware.

• Minimal actuation. The heads for BracketBot and Nori seem likely to be completely fixed.

This means they share a common set of limitations - no stairs, cannot access top shelves in a kitchen. Given the added complexity of legs, though, and the safety risks of a robot autonomously going up and down stairs, it’s not clear this is such a bad thing.

All in all — an exciting set of announcements and I would love to try out all three.

Leave a comment

Share

Subscribe now

It’s really hard to tell which robot policies are better than others, and who is actually making the most progress towards embodied general intelligence. I’ve written about robotics evaluation many times before, but this time I wanted to talk about leaderboards, as inspired by a thread on X.

There are many benchmarks, but few evolving leaderboards. PolarRiS, for example, is a great tool for benchmarking which correlates well with real-world performance, but it isn’t a live leaderboard in the same way — the authors do not update it, and the set of entrants is closed to the ones they chose at paper writing time.

Instead, let’s look at what public artifacts we have over time, from people who have been updating results over time, and see if we can identify any trends in robot learning.

Updated June 25, 2026 to add new Cosmos 3 results.

Subscribe now

Why This is a Hard Problem

There are two challenges here:

• Running a benchmark is expensive

• You need buy-in, which means community excitement as well as a feeling of progress (the benchmark must be hard but not too hard; real but not too real)

I actually briefly ran a robotics competition for open-vocabulary mobile manipulation (OVMM). There are a few problems here which make this uniquely hard for robotics compared to computer vision leaderboards (even something like the chatbot arena).

First, it’s costly to run a leaderboard like this, especially if your main audience is academics. Unlike LLMs, robot evaluation is costly: you need either large-scale simulation, or actual robot hardware, neither of which is cheap. Robot hardware in the real world will also incur resetting costs which do not exist for a language or image-based benchmark.

Another issue we’ll have to contend with: once a benchmark becomes widely accepted, it loses some of its value because people start to cheat.

Benchmark Hacking

Another big issue is benchmark hacking. We have seen this before with major LLMs like Meta’s Llama 4, but it’s rarer in robotics, partly because our benchmarks aren’t of widely accepted value like Chatbot Arena is/was.

This appears to have changed:

Read the thread from Pranav Atreya on X for more info. RoboArena works by volunteers evaluating pairs of policies; some teams were volunteering to assess policies, but often not completing all the evaluations they were assigned.

You’ll notice that now they’ve added indicators to their leaderboard for which models are low-sample:

Spirit, currently vying for top spot but with low confidence, is a fast-growing Chinese AI startup. Their VLA (Vision Language Action model - basically a generalist robot policy) is in fact open source, and you can find it on Github. For a primer on VLAs, check out my previous blog post on the topic.

My Own Attempt: The HomeRobot Benchmark

When I was at Meta we wanted to build a real-world robotics benchmark for pick and place. This is one of my favorite tasks still, and one that I think the robotics community remains fairly bad at because dexterity and task-specific work has overtaken it in focus.

Here’s the problem: I want a robot which can help me out in my home. That means I need to be able to tell it to pick up any object, and move it to any other location. As a hint we also gave it an idea of where the object could be found (sofa, chair, etc), as with low-resolution cameras maybe you cannot see a small object from across the room, and in a large house this will increase problem size tremendously.

In our case, we split this between a simulation-only leaderboard, exploring a bunch of different simulated homes (in the Habitat simulation), and a real-world setting in a specific apartment built by Meta for the purpose.

The simulation component was implemented on EvalAI. Basically: submit a docker image containing your code, it gets executed by us, and we maintain full control of the test environment. This is a bit different from most modern leaderboards, which emphatically do not allow for on-site code execution — but of course it also dramatically affects hosting costs.

In the end, the experiment was a middling success; roughly 60 teams submitted, we briefly evaluated three of these in the real world. We learned a lot though.

It’s very hard to strike a balance between “too hard” and “too easy.” OVMM was a case that was actually pretty hard for learning-based methods, and in fact our winning solution in sim and real was a model-based method that merely leveraged open-vocabulary vision.

A lot of the time was spent chasing down simulator bugs or tweaking RL environments to make sure they trained properly — in order to give those using the environments something to actually work with, and make sure the environment actually transferred to the real world.

Another note is that the $25,000 Hello Robot Stretch was probably still too expensive. Today I would probably want to use something like XLERobot or BracketBot — at less than 1/10 the price, everyone could buy one and use it. This would also make it more important for us to set up “home testing” instructions, codifying the rules we could use so people could scatter objects around their own labs to get a good signal about how well they would do.

Three Leaderboards

But enough about that. Mobile robotics benchmarks are particularly difficult, as they just take up so much time, to build and to evaluate. Most of the excitement today is around robot dexterity — robots manipulating objects.

Let’s look at three interesting leaderboards for robotics models that have all been updated recently:

• RoboLab-120: a simulation-based benchmark with experiments run by NVIDIA staff (leaderboard, code)

• RoboArena: a distributed, simulation-based robotics leaderboard where experiments are run in reality (leaderboard, RoboPapers episode, code)

• MolmoSpaces: an open-source platform for robotics develeopment, supporting multiple simulations and training methods. Technically this one is a mobile robot leaderboard, with a range of different tasks (leaderboard, Robopapers episode, code)

What do we see when we compare them?

RoboLab-120

Recently, RoboLab-120’s authors added Cosmos3-Nano-Policy to the leaderboards. This pushed down pi-0.5 and DreamZero, previously the top two previous policies.

RoboLab is a procedurally generated environment simulation based on pick-and-place tasks. Most of its benchmark entrants are VLAs. You can check out the list of example tasks here: they are dominated by pick and place, all in simulation, and run in different environments.

Some examples:

The environment leaderboard is dominated largely by VLAs, and all evaluations are run in-house. This is in many ways the gold standard for a benchmark, because you can’t just submit your own run (impossible to game, as happened briefly to RoboArena), but also means that the number of individual models is very limited — no Spirit VLA, for example.

The breakdown is interesting too. We see the newest NVIDIA model at the top, perhaps unsurprising for an in-house NVIDIA benchmark. The old pi-0.5 is second place, unless you rank for complex tasks only — in which case DreamZero comes in second.

RoboArena

We can see several world action models at the top (DreamZero for example), and the Spirit VLA mentioned above. There’s a much wider range of policies here, though the “confidence” can be a lot lower - there are so many more policies being submitted, and so many fewer potential evaluators.

If you have not seen it, this image from the website is a good indicator of what the policy evaluation environment looks like:

Basically, it is a benchmark of a single Franka arm performing various tasks, again largely pick and place. It’s another leaderboard that is completely dominated by VLAs and World Action Models, with Spirit,

MolmoSpaces

This is probably the most different experiment. MolmoSpaces is large and can involve mobile robots. While AI2 provides data, it’s not going to be as well studied as the Droid dataset. Correspondingly, we see a very different set of benchmark results here — some hybrid methods, including TiPToP, a task and motion planning-based system that leverages foundation models.

The wider range of tasks makes this benchmark more interesting, but it’s a pure simulation benchmark with lower-quality rendering.

Update June 25, 2026: Cosmos 3 is now at the top of this benchmark as well!

It’s definitely interesting how different these various leaderboards still are, with TiPToP (a task and motion planner) near the top of MolmoSpaces, Cosmos crushing the simulation benchmarks, and a list of VLAs on RoboArena.

Final Thoughts

There are many other benchmarks out there. I’ve written at length about evaluating robot policies before, and won’t go into that in more detail:

But let’s look at what the top entries here actually are. When it comes to pick and place tasks, there’s not a strong consensus — the ordering of the top models seems in flux. Compared to LLMs, though, it seems older models (especially pi-0.5) have a lot more staying power. It’s also just possible that benchmarks like RoboArena are easier to saturate - that we need complex, and more challenging, benchmarks with associated resources for verification.

So, my takeaway: we still really don’t know what is the best here, and there are no clear winners who just dominate all available benchmarks. Anyone could still “win” in robotics. We see new players (Spirit) appearing at the top of established leaderboards; we also see a surprising stickiness for some older policies.

Part of that last stickiness is almost certainly data — pi-0.5 was trained on tens of thousands of hours of high-quality real robot data, which many newer entrants have not yet managed to collect. This will put them ahead, at least until the large data collection industry can be leveraged to bolster their performance.

Let me know what you think; and what you think the future robotics leaderboards should actually look like.

Subscribe now

Share

Leave a comment

Robots need to be able to interact with the world in order to be useful. They do this in many ways: by driving around and taking pictures during inspection tasks, by walking up and down stairs. Military robots often interact with the world by exploding. But increasingly, we want useful robots to manipulate their environment and perform human-level tasks, which requires a gripper or hand.

The problem is that human hands are incredibly capable — they’re a high bar. Human hands are strong, light, adaptive, and self-healing. The human fingertip has ~2,500 touch receptors per square centimeter, with roughly 22 degrees of freedom. It can move at high speeds while applying significant force.

Building a robot hand that can match human capabilities, then, is very difficult. But robot hands have seen incredibly rapid change over the last couple years. You might have seen incredible manipulation demos from companies like Sharpa Robotics. The Sharpa Wave hand is part of a new generation of robot hands which nearly match human capabilities.

Now we have hands like this one, which is $600 and precise and compliant enough to pick up a raspberry without smashing it:

This is a great task for a few reasons: a raspberry is obviously very tiny, but it’s also squishy, soft and easily damaged. The CRAFT hand was made by researchers at UIUC and UC Irvine, led by Prof. Unnat Jain. The price tag is incredible, at only $600, and the plans for building it are open source. But there are a wide range of open-source robot hands coming out now, as well as an increasing range of commercial options.

So, let’s talk about robot hands. What are the key design decisions? Which one should you get? How do you make sense of all this? If you find this interesting, please subscribe and share.

Subscribe now

Why Do We Even Need Hands?

The most common robot hands, like those mounted on the YAM or ALOHA arms, have only a single degree of freedom: opening and closing. And this is enough for an incredibly wide range of tasks, as Physical Intelligence is so adept at demonstrating. Obviously we do not need hands on our robots then:

But of course, no one sees a demo of something that doesn’t work; robots aren’t deployed doing things that cannot be automated. And indeed it turns out that lots of tasks can’t be done without multiple fingers: in-hand reorientation of an object, for example. And solidly grasping long, thin objects — even managing door handles — can be difficult.

While the limitations of simple two-finger grippers are not an issue in relatively-controlled industrial settings, and for simple pick-and-place tasks of small, low-mass objects, this describes a minority of tasks that we might care about. Things that we need higher-DOF dexterous hands to do include:

• Very stable manipulation of objects with a wide range of masses

• Robust handling of long or thin objects

• Tool use, especially for any tool that is designed to also support humans

For a truly general-purpose robot, robot hands need to get a lot better. But this itself is a huge issue; human hands have a frankly incredible mix of robustness, ability to apply force, sensor density, and dexterity, with degrees of freedom comparable to entire humanoid robot platforms.

Fortunately, many people are working on this.

Types of Robot Hands

Robot hand design has to trade off between many different conditions:

• Strength - how much force it can apply

• Compactness - how big the hand is — human size being ideal for video learning

• Precision - how accurately can it servo to different positions

• Affordability - how expensive is it?

• Robustness - how often does it break

Human hands can apply a huge amount of force in a very small and very durable package, and two of them are attached essentially to every human. They therefore max out nearly every one of these conditions. Robot hands lag behind a bit. To understand why, we need to talk about how robots actually actuate and interact with the world.

A large part of the modern “humanoid revolution” and the scalability of robot quadrupeds is the use of quasi-direct-drive (QDD) actuation pioneered by the MIT Mini Cheetah. Quasi-direct-drive here usually means a low-gear ratio planetary gear assembly, connected “directly” to the joint that it’s moving.

This gives a great tradeoff between force transparency (equivalence between applied and measured forces), and backdriveability (ability to “push” a joint backwards without exerting extra force).

These two qualities are extremely important for robot hands, which are constantly colliding with things, and thus experiencing forces all the time! These forces can easily translate to robot damage, and so a non-backdriveable hand with poor force transparency can be easily destroyed by simple interactions.

The problem is that putting the big motors that you need in order to make a robot work are not (normally) going to fit in a human hand as they would for the leg of a quadruped or the arm of a humanoid robot. To exert the forces you want to firmly grasp and hold objects, then, you need higher gear ratios, which hurts backdriveability and has downstream effects on performance. In addition, these servos also mean more weight in the fingers, which causes its own performance problems and further increases the need for powerful servos further up the kinematic chain.

Broadly, there are two classes of solutions to the problem:

• Tendon-driven hands, like the ORCA hand, put the motors outside the fingers and connect them to joints via a tendon of some kind.

• Direct-drive hands like those from market leaders like Wuji and Sharpa, whose strategy revolves around building smaller and smaller motors to actually fit into the individual joints of the hand, using a higher gear ratio plus active compliance to achieve backdrivability.

This blog post by Quanting Xie, Tongzhou Liao, and Yonatan Bisk called the Dexterity Deadlock is a good read for more on the topic.

And there’s another connected issue. Once we are starting to pack lots of actuating power into a small form factor, we might need to make trade-offs and reduce the number of motors. This cuts the number of active (actuated) degrees of freedom of the hand, but results in lower cost, weight, and can save size as well. Notably, even human hands have passive degrees of freedom (the top joint on your finger usually cannot move independently of the others) — even Mother Nature struggles to produce a perfect hand.

It’s also possible that the actuation problem above needs to be solved by wholly new technology. One example of a team trying to do this is Daxo Robotics, which is working on soft or continuum robot hands which lack rigid links and thus technically have “infinite” degrees of freedom, similar to human muscles. These technologies are currently harder to manufacture, but this might be solvable; and they can take a lot more punishment than many traditional hands. For more you can check out our podcast episode from last year:

RoboPapers

Ep#40: Daxo Robotics

How can we build robotic hands with truly superhuman dexterity? Daxo Robotics is developing a unique tendon-driven soft robot hand, which aims to be tougher and more capable than a traditional humanoid hand. Each finger consists of many different tendons, which act in concert to move or manipulate…

Listen now

8 months ago · Chris Paxton and Michael Cho

Open Source Hands

We have a very large design space of tradeoffs in robustness, strength, cost, weight, and flexibility to explore, in order to try to build robot hands that can get close to human level performance. Much of the research I’ve seen looks at tendon-driven hands: academic researchers aren’t spinning their own motors or building custom high-precision gearing in the same way as an industrial hand manufacturer might be.

Another huge issue for researchers is the high cost of current hands: the best hands like the Sharpa Wave are often over $50,000 each, meaning that the whole cost of a robot can be dwarfed by the pair of hands mounted on the ends of its arms.

So there have been a couple recent efforts to build good, high-quality open-source hands. Take, for example, the RUKA hand from NYU, a <$1300 tendon-driven hand with 6 active and 15 total degrees of freedom for compliant grasping. It’s slightly bulkier than a human hand and has less degrees of freedom, but is lightweight and much more affordable.

Another example would be the LEAP hand (v1, v2); the v2 hand has 8 active dof, which makes it slightly more capable of adapting to different tasks and means it can be used for, for example, in-hand reorientation of objects. It’s much bigger than a human hand, being more similar to the Allegro hand in size, but still costs only ~$2000. The advanced version is even better, at $3000 and 22 dof, roughly human level.

The recent CRAFT hand actually starts to approach human capability on paper, with 15 active and 5 passive degrees of freedom, plus good compliance and actuation with lightweight fingers and for about $600. It’s a 3D printed hybrid hard-soft hand, which means that its finger joints use soft thermoplastic polyurethane, aiming to make sure that it can conform and make contact smoothly and robustly without taking damage from applying too much force to an object.

There are many more researchers working on novel types of robot hands; given how difficult the problem is, it is great that so many open, printable and hackable ideas are being produced to drive the area forward.

A New Wave of Hand-Centric Startups

And of course now there are a wave of robot hand startups appearing. This will not, by any means, be an exhaustive list of hands, just a few representative examples to talk about new entrants to the space over the last couple years. There are of course an older wave of hand companies like the Shadow Robot dexterous hand series — these too are out of scope for this post.

Some examples, but by no means an all-inclusive list:

• Sharpa: Singapore-based Sharpa is probably the industry leader at the moment, building hands which are a near match for human capability. They have done a number of impressive demos, including one where they show their 63-DOF Sharpa North robot peeling an apple with an apple peeler.

• Linkerbot is seeking a $6 billion valuation for a new round of funding (as per Reuters) and is mostly known for its robot hands. There are robot hands like the Linkerbot o6 which are roughly $1500, though this one is only 6 dof.

• BrainCo’s Revo 2 DuraX hands focus heavily on durability. The hands have been demonstrated many times as prosthetics to replace amputated limbs, similar to Psyonic hands. While the Revo 2 had only six active degrees of freedom and was relatively cheap, they’re moving towards higher-DOF and thus more expensive hands with the Revo 3, which will also integrate tactile sensing.

• The Psyonic Ability hand is another durable hand with only 6 active degrees of freedom, designed first for prosthetics and featuring built-in pressure sensors.

• Inspire hands are a commonly-used cheap robot hand, also with six active degrees of freedom.

There are many others of course. But the trend here is clearly towards higher degrees of freedom and more tactile density, with the SharpaWave being very similar to a human hand in terms of raw specifications: it has roughly 1,000 tactile pixels per fingertip and 22 active degrees of freedom, vs. roughly 21 for a human (depending on how you count).

What Issues Remain

The big open challenges in hands, as I see it, are:

• High cost. Robot hands are expensive. Two $50,000 USD SharpaWave hands may be the best on the market, but they will cost you more than your entire robot.

• Reliability. Human hands take constant impacts and bounce back, partly because they are ablative (covered in regrowing skin and muscle that heals on its own). Robot hands are notoriously finicky; even very robust hands like some of the prosthetics mentioned above won’t match human reliability.

• Sensing. Part of why human hands are robust is also just that we use them better: we avoid damage to our hands, we respond very quickly to heat or pressure by instinct. This is served by a network of dense tactile sensors and built-in whole body control behaviors centered around preventing damage.

None of these issues, I think, are insurmountable, but all will take dedicated work. There’s no “perfect” hand on the market now, which is part of why so many research labs have decided to build their own.

Final Thoughts

This is a blog post without a real lesson, just an observation. There’s a lot of companies and researchers out there trying to solve the problem of human-level dexterity. Now, for the first time, there really appear to be roughly human equivalent hands, in terms of dexterity and sensing if not manufacturability, strength, and robustness.

But these remaining issues will be resolved. This has real implications; as Avik De and I wrote about in our recent article on the changing nature of robotics research, hand researchers will shortly need to “move up the value chain” away from the kinds of fundamental work they are doing now.

min{power}

The First Paradigm in Robotics & AI Research: Lessons from Computer Engineering

Read more

2 months ago · 29 likes · 8 comments · Avik De and Chris Paxton

There’s no version of the world, I think, where in 5 years there are not good, dexterous, capable hands for a few thousand dollars price, ready for integration into different robots and products.

You should probably plan accordingly.

Subscribe now

Share

Leave a comment

References

• Lin, L., Patel, S., Moon, J., Lazebnik, S., & Jain, U. (2026). CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance. arXiv:2603.12120. Project page: https://craft-hand.github.io/

• Zorin, A., Guzey, I., Yan, B., Iyer, A., Kondrich, L., Bhattasali, N. X., & Pinto, L. (2025). RUKA: Rethinking the Design of Humanoid Hands with Learning. Robotics: Science and Systems (RSS). Project Page: ruka-hand.github.io

• Xie, Q., Liao, T., & Bisk, Y. (2026, February 12). The Dexterity Deadlock. Origami Robotics. Article: origami-robotics.com/blog/dexterity-deadlocks

• Shaw, K., Agarwal, A., & Pathak, D. (2023). LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning. Robotics: Science and Systems (RSS). Project Page: leaphand.com

• LEAP Hand v2 Advanced: Dexterous, Low-cost Hybrid Rigid-Soft Hand for Robot Learning. (2024/2026). Project Pages: v2.leaphand.com / v2-adv.leaphand.com

For a long time, we have seen robots mostly in ones and twos in videos, mostly in lab environments. But there have been a spate of announcements this week that aim to change that, with evidence of promising real-world deployments and real manufacturing scale-ups starting.

A quick reaction blog post, and some thoughts. If you find this kind of thing interesting, please consider subscribing.

Subscribe now

RobotEra Logistics Deployment

Check out this post from CyberRoboooo on X, which shows a line of RobotEra L7 humanoids in action. These robots are busy aligning bar codes at logistics hubs all over China, including ones run by by China Post and others.

Aligning bar codes looks like a ridiculous task for humanoids, but as you can see here it’s actually one that’s rarely automated. This is because:

• It requires handling a surprising variety of package shapes and sizes

• Requires human-level dexterity and speed — traditional automation struggles

• Requires high reliability - every bar code NEEDS to be scanned.

At really high-end logistics companies (think Amazon), this is indeed done with traditional automation — packages go through a funnel with many cameras, so orientation does not matter. But these solutions are expensive, and most e-commerce retailers don’t use them yet, instead relying on human labor. Which may become humanoid robot labor soon.

Beijing-based RobotEra launched this logistics platform last year. RobotEra has raised around $200 million USD equivalent funding (RMB 1 billion) and builds a wide range of robots.

A Cheap Unitree Manipulator

Unitree is most famous for its G1 and H1/H2 humanoids: mobile robots which are largely not designed for manipulation. Sure, they’ve been used that way — many times — but they’re fundamentally robots for whole-body control research and entertainment. While they’ve toyed with manipulation-focused robots in the past (like the Unitree G1-D), this hasn’t really been an area of focus for the company.

So it’s really interesting that they’re launching a line of manipulation-focused robots based on their very cheap R1 series platform. The R1 is the unitree “toy” robot, the roughly $5k humanoid you can buy to cartwheel and box and, well, that’s about it, normally.

With 19 degrees of freedom (including 2 from the grippers), a mobile base, an actuated head, and ~half a meter of reach, this actually seems like a really nice little robot to get started with, especially given the low price. The payload is very low, at only ~2 kg per arm, but is still enough for lots of light manipulation work.

Information about the robots was posted on X. The price doesn’t seem to include the mobile base, but it’s still a very good start and indicates how cheap, at the low end, industrial semi-humanoids are likely to become.

Figure BotQ Manufacturing Update

Not to be outdone by the Chinese companies, a few American companies have made huge announcements with incredible high-quality videos this week. First up is Figure, which increased production of their Figure 03 humanoid up to 1 robot per hour — capable of 24 a day.

Stills from the videos in their blog post are impressive: showing a veritable army of robots, robots performing exercises, human technicians building robots, and perception-aware whole body control.

Another interesting note: Figure is now doing 3d perception. Their RGB stereo cameras are used to predict 3d structure, which they can use for navigation and moving around in the environment. Another step toward real-world deployment, as it means reliable, robust navigation should be ready to ship soon.

I’ve written about Figure 03 before if you want to know more about the robot. Currently, these robots are targeted at research, data collection, and customer discovery.

1x Hayward Factory

Home robot maker 1x released a look at their Hayward factory, built to produce 10,000 NEOs a year. The look even showed NEO robots walking around the factory floor, helping out with various tasks.

Some more points they announced:

• Vertical integration means building motors from copper on up

• 10,000+ pre-orders in the first 5 days of their NEO launch (order here)

I won’t say as much here because the videos speak for themselves. You can also read the accompanying blog post.

Tutor Intelligence and “America’s First Data Factory”

I wanted to include this: Tutor’s co-founder and CEO Josh Gruenstein shared a 100-robot data factory, where their industrial semi-humanoid Sonny is busy being used to scale robot learning.

Data Factory 1 (as they call it) aims to scale remote-VR proprioceptive teleop, so they can use it to cover the whole range of the robot’s lifecycle. This includes data annotation:

This is a nice look at what a good teleop system looks like and how to annotate images; a “peek behind the curtain” of scaling robot learning that many may not have seen before. These operational details are maybe not useful for Tutor’s competitors (the Physical Intelligences of the world) but are really useful for those getting started, so it’s great to see.

Final Thoughts

While manufacturing progress is being made, much of what is being scaled now seems to be “industrial semi-humanoids,” i.e. integrated robots with multiple arms and a head capable of a wide range of dexterous tasks. RobotEra, for example, does not bother with legs for their package-handling deployment. You’ll note that the RobotEra task - package arrangement — is very similar to one Figure has been doing, and Figure is also perhaps unsurprisingly in talks with UPS to deploy their robots.

I think we will see that there are a few tasks that modern end-to-end robot learning is extremely suited for — box packing, package sorting and reorienting, clothes folding — and that will proliferate quickly. The costs of these activities will rapidly drop as robotics production and intelligence scale.

Subscribe now

Share

Leave a comment

Large language models are a fundamentally different type of intelligence from humans; and this means that even as they approach and even exceed humans in some domains (coding!) we will still see some places where they lag behind human intelligence in surprising ways.

Spatial reasoning is one of these — telling the difference between “left” and “right” in various contexts is still often hard for even very powerful models. Says one user on X:

This shortcoming of language models has actually existed for a very long time and it’s an interesting one that, in my opinion, is illustrative of the structural weaknesses of learning principally from language, and why these weaknesses aren’t as true for, say, color. And so, a short post on left, right, colors, and why some things are harder for LLMs and VLMs to learn than others.

Subscribe now

Spatial Understanding is Hard from Language

Language does a terrible job specifically of capturing spatial information. Why is this?

Spatial concepts just aren’t obvious and well-defined from the vast corpus of text available to train on. What does it really mean to walk across a room, or to put a book on a bookshelf? The precise spatial locations that we might need to unambiguously resolve this missing information into a clear concept just isn’t a part of the training data in any real way.

I’ve written about world representations for robots in the past (a term which I use very differently from “world models”). For my world representation to be useful, it needs to capture, for example, the positions of all of the various objects in a house, so that I can navigate to them, manipulate them, and ask questions about them.

The process of using sensor measurements to answer these questions is usually referred to as grounding the concepts. The symbol grounding problem deals with how we connect these symbols (on top of a table, inside a box, lined up, clean…) into an understanding of the physical world.

There’s actually a surprising amount of information in LLMs on how objects should be arranged, but there’s always missing information too — details about geometry and contact, about object shapes and physics. Training on text is like learning from a textbook with half the pages torn out, or watching a YouTube video with captions in the wrong language. You’re probably still going to learn something, but a lot of the key context is missing.

This Missing Information Persists in VLMs

But many of the leading frontier models these days are actually multi-modal, taking in images or even short video clips. Unfortunately, this is still not enough to resolve the problem — it persists even in a vision-language model (VLM).

VLMs are trained on images from a single perspective, or in the case of a short video a handful of highly correlated perspectives. This reduces ambiguity and helps resolve some of the missing context, but hardly removes it completely. What does “left” mean in the context of the portrait? My left, or the left of the person in the picture facing me? These things are opposites.

Such ambiguities aren’t really a challenge for modern VLMs (I can just specify “my left” or “their left” in the prompt), but complexity and ambiguity remain. Problems like precise estimation of 3d object position from images remain challenging, which means that for many real-world tasks, the VLM just doesn’t work to a reasonable level of reliability.

Example: Position vs. Color

Consider how an LLM might leverage its training data to understand colors (red, purple, blue..) versus left and right.

Color is used consistently. An apple is red; the sky is blue; leaves are green. Even if I do not know what green “means” — in terms of specific RGB values — I do know which things are described as green, and because green represents a distribution of RGB values, it’s used fairly consistently and it’s easy enough for a model to understand that something will be green just based on this data.

However, as mentioned above, left and right are highly context dependent. They’re going to be used in millions of different ways, to mean different things.

Broader Implications

This is an important illustration of an idea that I’d like to write more about: I fully expect AI and robotics, now, to be able to exceed humans in many domains. This does not necessarily mean they’ll exceed in all domains; there will be some “comparative advantage” where humans remain leaders in some tasks. It’s hard to say what those are; but we can look at the types of learning and reasoning that humans do well for examples.

Humans, for example, operate off of comparatively noisy training data. We set our own learning goals rather than relying on someone else evaluating us. We get bored and do something else, managing our own time to keep us from getting stuck in pointless loops (except during grad school).

Our “training data” is based towards real, streaming, multi-modal data that’s aggressively compressed - using far more compression than works with modern transformers. Humans have to disregard millions of useless signals from various sensors just to go about our daily lives. By contrast, LLMs and VLMs are trained using mostly still images and mountains of text — they’re learning concepts like “left” and “right,” “purple” and “green” all from associations.

There are lots of cases where machine and human intelligence just end up working differently, due to these differences in compression, training data, and of course our different computational hardware.

Ethan Mollick has written about us approaching a sort of jagged AGI, where modern models exceed human intelligence in some areas but fall behind in others, making comparatively little progress. When you look at how these models are learning and what they are doing, I think we’ll find many of these shortcomings are due to fundamental differences in the way human vs machine intelligence works.

This is obscured by the fact that usually, we don’t need our AI agents to interact with the physical world. They inhabit a realm of text prediction and function calls just fine. But this still represents a fundamental limit of what they can do, which can only be overcome, I think, by figuring out how to scale embodied data.

Final Thoughts

This is all to say that certain types of intelligence, I believe, only really make sense with embodiment. Text, and even much of the image data on the internet, fundamentally can’t learn to solve all of artificial intelligence — embodiment matters.

This is also another reason to be excited about world models, which potentially can start to understand these concepts and learn to ground the various spatial relationships that we talk about in reality. World models take in and predict video data; because of this, they can actually learn an implicit model of the world, of how objects interact, and learn (implicitly) to understand the 3D space that humans and embodied agents must inhabit.

Perhaps the most interesting question, then, is this: will embodied data eventually allow our AI agents to become better at, say, using 3D modeling software, as per the tweet in the intro to this blog post? I believe so; it seems possible at least that these concepts will be best captured through large scale interaction.

After all, it would make sense if, to achieve human-level artificial intelligence, we needed agents to learn the same way humans do.

A horde of visitors just descended on San Jose, California; blocking off multiple city blocks, many wearing black and lime green, all of them talking unceasingly about artificial intelligence. Yes, it’s time for another NVIDIA GTC.

GTC is an interesting conference for me. It’s much more focused on business and sales than the conferences that I usually would go to, like the Conference on Robot Learning. Most of this is selling AI software, but there was a pretty good presence centered around physical and industrial AI — which is to say, robotics.

In this short post I’ll give some of the things I saw on the show floor.

All videos and photos in the article were taken by me.

Subscribe now

Skild Assembles GPUs

First up, demoing robotics manipulation in the NVIDIA booth, it’s Skild!

Skild has just raised an incredible $1.4 billion Series C with $30 million in revenue, meaning it probably has the most revenue of any of the current-gen robotics “neolabs.” And here they’re demonstrating why: they have an incredibly precise manipulation demo showing server rack assembly.

From CEO Deepak Pathak on X:

• This is showing fully end-to-end learning on a real industrial task

• It’s just using cameras - no fancy, expensive sensors with arduous calibration procedures

• Long-horizon reasoning and action via in-context learning

• Robust to disturbances and easy to deploy for new tasks

According to the booth staff, they had 15-20 hours of real robot data for this demo, as well as use of NVIDIA Isaac Lab with Newton; and in fact we can see the task simulation in the NVIDIA promo videos for Newton.

Noble Machines Shows End to End Humanoids

Noble Machines just recently came out of stealth showing their brand-new industrial humanoid.

They showed demonstrations of whole body control for robot teleoperation, as well as a couple autonomous manipulation policies: picking and placing 2x4’s and pipes, and loading and moving a tote. Both worked flawlessly at least while I was watching.

I also like how this robot is wearing shoes; it might hurt motion aesthetics (because it’s unlikely to be modeled properly in Isaac) but it’s an easy solution to the incredible difficulty of some shop and real-world environments. Shoes are “ablative” and easy to replace, and they have a lot of grip on the floor by design.

Watch some more video of it in action:

Generalist Shows Incredible Dexterity

Robotics learning startup Generalist has quietly been building up impressive technology. In the Universal Robots booth, they showed their manipulation capabilities live. The task: putting a phone in its box. Some impressive details:

• They only had the robots for a few days to prepare for the demo - this was a bit last minute

• They claimed mere hours of on-embodiment data - a very fast setup process, though they likely had quite a but more data from the same task on other robots

They were constantly doing things like hitting the robots with hockey sticks, interrupting the manipulation and watching it recover.

Generalist has since gone on a bit of a press tour, so I will leave further details to them.

Agile Humanoids

Agile showed both humanoid robots performing a range of manipulation tasks, and a teleoperation system. Agile Robots is one of the small but growing number of European companies in the space, showing how the robotics race is going global.

The Unitree H2, In Person

My first time seeing this robot in person. I think this will be huge — likely the biggest humanoid robot in the world in 1-2 years.

You can also see it fighting at GTC — check it out on X.

Pony AI Car

Self-driving cars continue to scale up. I won’t say much about this, but you can check out my previous blog posts on their competitors like Baidu and Waymo.

Medical Robotics

We don’t see as much work on medical robotics making waves on social media. Medical robots don’t make for flashy videos, but they’re extremely important

Cheap Robots

Seeed studio builds a variety of open-source and low-cost robotics hardware, powered by NVIDIA Jetson. It’s interesting to see this here, as GTC is mostly aimed at larger businesses.

They are preparing to launch re-Bot, which is a larger, higher-payload arm more similar to the common YAM arm. The availability of extremely cheap open-source hardware like this will make robotics research and experimentation far more accessible, which is a bright sign for the future.

Olaf Walks The Floor

NVIDIA and Disney’s robotic Olaf the Snowman looked fantastic, like a Disney character come to life.

Interactive Humanoid Demo

Humanoid had an interactive demo running in the exhibit hall, where you could speak into a microphone and ask the robot to bring you something. This company has been moving incredibly fast to get their robots up and running and start early deployments.

Dexmate Robot Wanders the Hall

Dexmate didn’t have a specific booth, but they still brought a robot. The Santa Clara-based company is one of relatively few aiming to build a platform, with high-quality general-purpose hardware that people can build applications on top of. The Dexmate Vega itself looks highly capable, and was safe enough for them to take it on the show floor.

I liked seeing the teleop rig they were using; it’s easy to imagine a version of this being used to do mobile manipulation tasks in different locations.

Final Thoughts

Compared to previous years, we saw:

• Far more end-to-end demos

• More mobile manipulation

• More teleop

Just more robots that appeared to work, in general. It’s also striking how these companies are all aggressively pursuing applications now; this did not feel like a crowd of researchers discussing their latest papers. Everything felt very substantial and tangible.

All in all, an exciting look at the current cutting edge in robotics.

Leave a comment

Share

Subscribe now

Even if we should be freaking out about AI, intelligence specifically in robots still has a long way to go — it can take weeks to deploy a robot for a new task. On the other hand, video generation has gotten insanely good recently: Veo 3 from Google and Seedance 2.0 from ByteDance seem likely to change the way Hollywood works forever.

Perhaps as a result of this incredible jump in capabilities, world models — which is to say, video-predictive model for robot tasks — are one of the most exciting trends in robotics right now. They show advanced generalization capabilities, can execute open-ended tasks, and provide an effective way to leverage large amounts of internet data to learn “priors” about how the robot should interact with the world, thus providing an effective way to close the robot data gap.

I’ve written about world models before, so think of this post as an update — looking at the developments over the last few months and a couple novel ways of employing these models, as well as some trends that have developed.

If you like this blog or these kinds of posts, please like, subscribe, and share:

Types of World Models

I’ll go over a few different types of world models that people look at. Note that I am to some extent using my own terminology here; terms like “world model” and “world action model” get used all the time for all sorts of things, so I’m going to give you three different ways of setting up the problem. There are, of course, tons of types and subtypes within each.

Action-conditioned World Models, like V-JEPA 2 [1], which predict future states based on an action vector. These are the most theoretically justified, and are the “model” in “model-based reinforcement learning” — a very well-studied but still nascent area of robot learning.

These (very, very roughly) take the form:

where f is the world model, x are state observations, a are actions, and x’ is the predicted future state sequence (image or video). They are the “classic” formulation of the problem, and have been famously championed by Yann LeCun. Other great examples are Dreamer v4 [2], which learned a “simulation” of minecraft and was able to train policies within that world model; and RISE [3], which is a recent paper that learns a “local” action-conditioned world model and uses it for reinforcement learning.

Next, there are Video World Models, which work by generating a video first, and then work backwards using an “inverse dynamics model” to recover robot actions — DreamGen from the NVIDIA GEAR Lab is an example [4], as is the most recent 1x World Model, and the LingBot VA model from RobbyAnt [5].

These have the advantage of most directly being able to leverage large-scale video data, as they do not need actions at all to train the world model, but instead recover actions from this “inverse dynamics model” g which can be trained on far less data:

Finally, joint World Action Models (WAMs) are policies which predict both video and action jointly -- DreamZero is an example [6], as is Fast WAM [7]. While the term has been used many different ways, I’ll use it here to mean a joint prediction, indicating that it learns both how the world evolves and what actions the robot is taking together:

These models seem to generalize better than alternatives, as explored in the paper “Do World Action Models Generalize Better Than VLAs?” by Zhang et al [8].

Separately, there is also the question of 2D vs. 3D. There are also 3D generative world models - which create a 3D scene, such as PointWorld [9] or Marble from World Labs. This is very broad category, deserving of its own post in the future.

Background: Planning For Robots, or: Why Do We Care About This?

Planning for robots is a field with a long history, dating back over 50 years. People have been using robots to solve complex, long-horizon problems forever. The issue is that these always rely on models of the world, which have historically always been defined by humans.

The key attribute of any of these models is the dynamics model, or how the world state x will evolve in response to the robot’s actions a. From this, all else follows: we can use a good dynamics model to plan over complex terrain and solve challenging symbol manipulation problems.

More recently, large language models have proven themselves to be reasonably good at this kind of thinking, but they remain unable to solve the symbol grounding problem which stymied the old-school planners: basically, what does it mean to “grasp a cup by the handle,” or “be careful, that’s hot,” or really convert symbolic instructions into measurable real-world values.

Fortunately, we have near infinite amounts of video data available to us in order to solve this symbol grounding problem now. And thus, world models.

Action-Conditioned World Models

Above: a screen recording of me using the Interactive World Models [11] online demo.

Action-conditioned world models predict a world state (usually a latent world state) given previous observations and robot actions. Examples include Interactive World Simulator [11], V-JEPA 2 [1], DREAMER v4 [2], and RISE [3]. Also notable is Genie 3 from Google, which generates scenes from a text prompt but then lets you explore them.

The main downside of these models is that they tend to be fairly local; errors accumulate rapidly, and they can’t do really long horizon reasoning. Even if they maintain physical consistency (which they increasingly do!), they will drift away from the real world.

I tend to think of this as the “purest” form of world model, and it is the oldest, despite its complexity and difficulty.

Alternate names:

• learned simulation

• JEPA-style model - technically a subset of action-conditioned world model which relies on specific tricks to prevent representation collapse when training on purely latent-state prediction without reconstruction

• learned dynamics model (as in the end it’s modeling the changes in world state)

Video World Models with Inverse Dynamics

But these action-conditioned world models have a crucial weakness: they fail utterly at long horizons. Prediction quality degrades rapidly. They tend to fail at “long” horizons, even to the order of seconds (though this is getting much better; again see Interactive World Simulator [11] and try the demo).

Another more subtle issue is that they require action labels. Action labels are not available for most video data: we do not know how a human’s finger joints moved to pick up the cup, only what the results looked like.

And so we have a class of models which can take advantage of the massive corpus of video data to predict robot actions. They are “ordinary” video models, trained or fine-tuned on robot data, coupled with an “inverse dynamics model” which takes in successive states and predicts robot actions from them.

I tend to think of NVIDIA’s DreamGen [4] as one of the first outstanding examples in this category, which used video models and an “inverse dynamics model” to extract actions from a fine-tuned version of WAN, though they don’t directly execute the actions but instead used them to generate neural trajectories for training.

LingBot-VA [5] is a cool bridge between these models and our next category. During inference, it interleaves prediction of video frames with an inverse dynamics model which extracts actions from them, making it a hybrid between a pure video world model and a joint world-action model.

Watch the RoboPapers episode on the 1x world model, one example of this category. We also have a podcast on DreamGen.

Joint World Action Models

At different times, people have used the term “world action models” to refer to many things, but in this paper I’ll use it as used in DreamZero to describe the class of “joint world-action models” which predict both state and action without explicit conditioning.

These are probably the current “best” models, with chart-topping results in a couple benchmarks like RoboArena. They have the advantages of both world models (being able to train on and benefit from massive amounts of video) and VLAs (straightforward architecture, direct prediction of actions without jumping through a lot of hoops).

As per above: the main advantage here appears to be that these models can learn from video how the world should evolve implicitly, while still mostly predicting actions. There’s evidence (see Fast WAM [7], above), that decoding video is not at all a necessary step; DreamZero likewise does not decode video at test time, only at train time.

Watch our RoboPapers episode on DreamZero.

Bringing It Together

DualWorld takes the unique approach of combining the “video generation” approach with the much more local JEPA model. It uses a WAN 2.2 5B video generation backbone to predict high-level actions, then uses a JEPA model for local control to track these video references.

This is a cool approach because you can inject, for example, tactile and force information into the more local, short-horizon JEPA model, while getting all of the generalization advantages of the large pretrained video model. Perhaps approaches like this could result in a world model version of “system 1, system 2” reasoning and solve the issues with both approaches.

You can check out their video results here.

Final Thoughts

There are a ton of world model - focused companies out there, all with their own particular version of what a world model “is.” Many of these are not robotics focused. Moonlake AI, for example, raised $28 million in late 2025 to build game worlds, though they have grander ambitions. Yann LeCun’s Advanced Machine Intelligence raised a staggering $1.03 billion for world models, initially targeting smart glasses like Meta’s RayBans.

A note of caution is that world models have, historically, really underperformed expectations. They’re not a new idea, and especially action-conditioned world models have always struggled at producing real value. What’s changed is these things:

• Models, especially video generation models, are much better

• More data is now available for training these models

• New architectures and improvements like video world models and the joint world action models described above

Interestingly, the disadvantages for all of these models are basically the same; as with many trendy methods like sim-to-real learning, they don’t yet achieve the high reliability that we see from on-robot reinforcement learning approaches like RL-100 or pi-0.6*. But this is likely to change, as this is an active area of research, and the models themselves seem so promising. More to come.

References

[1] Assran, M., et al. (2025). V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning. Meta FAIR.

[2] Hafner, D., et al. (2025). Training Agents Inside of Scalable World Models (DreamerV4).

[3] Yang, L., et al. (2026). RISE: Self-Improving Robot Policy with Compositional World Model.

[4] Jang, J., et al. (2025). DreamGen: Unlocking Generalization in Robot Learning through Neural Trajectories. NVIDIA GEAR Lab.

[5] Li, L., et al. (2026). Causal World Modeling for Robot Control (LingBot-VA). Robbyant.

[6] Ye, S., et al. (2026). DreamZero: World Action Models are Zero-shot Policies.

[7] Yuan, T., et al. (2026). Fast-WAM: Do World Action Models Need Test-time Future Imagination?

[8] Zhang, Q., et al. (2026). Do World Action Models Generalize Better than VLAs? A Robustness Study.

[9] Chao, Y.-W., et al. (2026). PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation.

[10] Fikes, R. E., & Nilsson, N. J. (1971). STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving. Artificial Intelligence, 2(3-4), 189-208.

[11] Wang, Y., et al. (2025). Interactive World Simulator for Robot Policy Training and Evaluation. Columbia University, Toyota Research Institute, Amazon, UIUC.

Citations produced by Gemini from the text of this blog and may have errors.

Learn More

A great writeup from the Moonlake team on world models on X

Interactive World Models Paper by Yixuan Wang et al.

Dual world: https://world-mind.github.io/DualWorld/

Some previous blog posts and podcasts from me that you can read or listen to:

RoboPapers

Ep#61: 1x World Model

Every home is different. That means that to build a useful home robot, we must be able to perform zero-shot manipulation of a wide range of tasks — which is a real challenge for robotics, since so many cutting-edge approaches require expert fine tuning on a small set of in-domain data…

Listen now

5 months ago · 4 likes · Chris Paxton

Data-first home robot company Sunday Robotics announced their series B raise for a unique and contrarian take on how to deploy in-home robots today. Their approach claims zero robot data, but still focuses on real-world-data only. This shouldn’t be a huge surprise; co-founder Cheng Chi was behind the game-changing Universal Manipulation Interface (UMI) paper. The idea is to use a tool, an exact duplicate of which is mounted on the robot arm, to collect all the robot data.

A short blog post here with some thoughts on what makes this company and their vision of how to build robotics unique. This is a “quick reaction” blog post, so apologies for any errors in advance. If you like my writing please like and subscribe.

Subscribe now

Why Zero Robot Data

Part of what makes the Sunday vision of robotics unique is their focus on glove-style data collection. Robot data is expensive, as their head of product says on X.

To collect real-world data, you need:

• Skilled operators

• Real robots, at real-robot prices

• Good teleoperation setups, which are often clones of the whole robot

• To be able to move this around into different environments to collect data

All of this is rather expensive. With tools, assuming this data is as good as real robot data, all you need is the first part - the skilled operators. This is something Sunday does; by skipping the latter three requirements, they can jump directly to paying for high-quality data at scale.

And Sunday pays very well for data, up to $60 per hour according to public job postings — likely more for other positions and for specific customer requirements.

The fact that tools are easier to carry around is also important. The key to scaling robotics learning is data diversity, something I have written about before. Tools let you scale data easily, because you can just walk around and use the tools for anything you want. There are a few Robopapers episodes with a similar premise you can listen to, if interested in this line of reasoning.

Despite how we’ve repeatedly seen that diverse scenes and tasks are important, it remains very hard to get a good distribution of data just with teleoperation data. While other approaches exist — like large-scale human video data — all of these approaches have tradeoffs. My (boring) personal take is that we’ll probably use many different data sources together, but I like to see that Sunday is committing to an extreme to see how far it can take them.

Rapid Hardware Iteration

So, how does this work in practice?

The Sunday Robotics approach seems to have centered on fast hardware iteration. Above you can see some images — shared on X — of how they iterated on the gripper. It’s a unique 3 finger design, a “mitt” which can be exactly mounted on the end of the robot. In addition, data collectors wear a “hat”, similar to the baseball hat worn by their robot, Memo, which captures wide field-of-view RGB camera images.

Their final glove is around $400, which means they can afford to have a distributed network of roughly 500 operators collecting high-quality, diverse data at massive scale.

In public comments, the company has discussed how much time and effort was spent just on data collection - they spent much of their time building the tools and collecting data, and comparatively little time training their models.

Finally, they constantly point out the importance of building a full-stack team: that iterating hardware, data, and training as a whole is crucial to deploying real-world general-purpose robots.

Final Thoughts

This is a very data-first version of robotics. They have taken a strong bet that good data at scale is crucial, and that the particular type of data that can be collected with UMI tools is ideal. It’s an interesting bet and a promising one; I’m interested to see how it goes and hope to see Memo out and about in the real world soon.

Leave a comment

Share

Subscribe now

Something strange has happened to software development over the last year: the machines have taken over. Software development is faster and easier than ever with the aid of tools like Claude Code and Cursor. Similar changes are coming in other industries; AI tool use will come for computer-aided design (CAD) tools next; it will make it easier to design a robot or other hardware in software, write the bridges to connect it to simulation, and deploy it.

This challenges a couple major assumptions that have been driving massive robotics investment over the last few years, so I think it’s worth exploring the idea.

Subscribe now

The Case for One Robot

It’s a cliche that you have the same iPhone and the same laptop as Elon Musk. Scale, right now, is king in software and hardware. Electronics products are impeccably engineered, optimized, and designed for manufacturing at scale.

One might assume the same logic would apply for robotics; in fact, one might assume it would be even more extreme for robotics. This is because robots need real-world data; robot models are data hungry, and there is a huge data gap to be overcome.

Co-designing robot sensors, hardware, and models is already very important for good real-world performance. This means that deploying one robot everywhere, an approach championed by Tesla and Figure, has a lot of appeal. You solve these hard problems once, then scale, scale, scale until the robot is affordable and incredibly capable. There are clear parallels to how humans “work,” as in how humans are employed in a wide variety of industries and to solve a wide variety of problems.

This is only sort of how humans work, though. Humans use tools. Incredibly inhuman automation is vastly more productive than humans for a wide variety of problems, from package sorting to harvesting crops. Humans provide the intelligence, but our bodies increasingly are not actually performing the labor — our minds are.

The humanoid-only champions have an answer here: that the robots will be using human tools. This is something that I am deeply skeptical of: if robots are doing all the work, why would you design tools to support legacy hardware (humans) forever?

Initial Conditions

First, let’s start with a few assumptions:

• Software development will become effectively “free” over the next year, meaning that you can use AI coding agents to produce any code you can clearly design, given about a month and a few thousand USD.

• Simulations will continue to improve, to the point where they’re operationally indistinguishable from reality, at least on a mechanical level, given some real-world data to “seed” them.

• Cross-embodiment learning — a “single brain” for all robots — works reasonably well, and it’s not hard to deliver intelligence for a new form factor.

• AI agents will improve their ability to use CAD tools for hardware and electrical design, and can “close the loop” with a simulation like NVIDIA Isaac, Mujoco, or potentially a new upstart like Lucky.

What would be possible in such a world?

The Robotic Cambrian Explosion

The Cambrian Explosion was a rapid evolutionary event, a period roughly 539-520 million years ago when a vast number of animal phyla appeared in the fossil record for the first time.

If these conditions exist — if agents can be trusted with hardware design, if they can be given reasonably-accurate simulations of real-world tasks, then instead of seeing a few humanoids appear everywhere, we would see a massive diversity of different robot form factors.

After all, the motors that power a robot dog and a humanoid and an industrial pick-and-place machine don’t need to be all that different. One might easily imagine a world where a robotics company builds modular components like LimX’s TRON 2, which can be recombined in many different ways according to the needs of a particular task. They might go farther, with a list of 4-5 supported actuators and a handful of common sensors, with shells and wiring harnesses designed on demand.

This would mean that instead of a single humanoid in a factory, you would see a family of dozens of interrelated robots; big cargo haulers, fast-moving inventory-checkers, loaders and packers, all build out of the same core parts and on the same hardware. If you want one for your home, maybe it’s using lower-powered actuators, and cheaper sensors, but again it’s still the same mind and the same parts.

Importantly, if cross-embodiment learning works — which, to an extent, it seems to — these robots are still all sharing data and still making each other more intelligent.

This to me seems like a much more “AI-first” way of viewing robotics hardware development, even if in many ways it’s still far fetched.

Why It Might Not Happen

A few potential issues:

Maybe contact-rich simulation remains intractable. Despite the best efforts of NVIDIA and World Labs and so on, perhaps contact physics never become quite good enough to fully automate the design process. This would make the economics much worse for our robotics Cambrian Explosion: you would need much more rigorous real-world testing, with more room for human engineering and intuition in the process.

Human data is more useful for humanoids. But we’ve actually seen from recent works that you can get incredibly good results just with in-domain data, and there’s also a large and growing body of work on cross-embodiment learning.

In-domain data might be too important. Basically, cross-embodiment learning might be good enough for a demo, but what if it’s never good enough for true, high-reliability cross platform performance? This would, again, make the economics of this Cambrian Explosion harder, since you’d need lots of real-world data for each one (if it can all be done in simulation, you’re fine).

And there’s some evidence for this one: in a recent blog post from Physical Intelligence, we see that they achieved their best results by mixing downstream customer data into the pretraining for their foundation model. Similarly, for the NVIDIA example above, cross-embodiment robot data may or may not have helped any more than other egocentric data.

This is a pattern we have actually seen many times; robotics problems tend to be these little isolated “islands” of data with less evidence of positive transfer between them than you would hope. On the other hand, this will almost certainly improve as we “fill in the gaps” with more robot deployments.

Final Thoughts

The truth, likely, will be a mix between the two: some general-purpose platforms, and a great deal more robots for a wide variety of tasks. For now, at least, real-world data seems irreplaceable, but this could all change fast, and it constantly is.

The real test, I think, will be if data-driven simulation really takes off and proves that it can handle complex contact dynamics. If this happens, it will be much easier to train specialized robots in simulation. This reduces the risk and means you can produce a much wider range of robots without all the expensive real-world deployments and data collection.

But please let me know what you think below.

Subscribe now

Share

Leave a comment

The team at NVIDIA recently released a set of robotics models; one that’s been driving a lot of discussion online lately has been DreamZero: a “World-Action Model” which predicts video in addition to robot actions. In particular, it achieved benchmark-topping results at two very interesting benchmarks this week: RoboArena and MolmoSpaces:

Let’s briefly go over what this means and what we might learn from these results. In this blog post I write:

• A very high level look at what makes DreamZero different from other policies and world models

• A brief look at RoboArena and MolmoSpaces

• Some speculation about what’s different and what might be causing performance gains

A lot of this was pulled from discussions on X. Upcoming content: we will have a RoboPapers podcast on DreamZero in the next few weeks, and next week on this blog I should have a more in-depth analysis of world models and how the landscape has changed in 2026.

What is DreamZero?

DreamZero is a “world-action model” from NVIDIA. In practice, this takes a lot of the concepts from world models — principally that video generation is useful for robotics — but changes the formula in a couple key ways. Most crucially, it jointly models action generation and video generation.

Usually, there are two classes of world models:

• Action-conditioned world models learn the mapping x’ = f(x, a), for state observations x and actions a. An example would be V-JEPA 2 or the world model from the recent RISE paper.

• “Inverse dynamics” world models like NVIDIA’s DreamGen or the 1x world model (RoboPapers podcast for DreamGen, RoboPapers podcast for 1x), which learn x’ = f(x) and then an inverse dynamics model a = g(x, x’)

This model is much more like a traditional robot policy, but also predicts future video - so we are learning something more like (x’, a) = f(x). All of these are a huge oversimplification of the true training objectives, of course — read the paper for details.

You can also compare to a traditional vision-language-action model: we are also predicting what the future will look like. This gives a much richer supervisory signal over how the world will evolve going forward.

The Benchmarks

RoboArena is a distributed “real world” benchmark based on Droid. Evaluators across the world have somewhat similar robots and setups, and run a set of open-ended robotics evaluations with different natural-language commands.

This means that it’s somewhat in-distribution for DreamZero, as it was trained on Droid, which contains very similar tasks and setups. But it’s still a real-world evaluation, with all the difficulties and changes that this implies; and the tasks are chosen by evaluators. It’s also a “head-to-head” comparison benchmark, similar to the Chatbot Arena which was so influential in LLM development. I wrote about this at length in my post on robot learning evaluation.

MolmoSpaces is a new benchmark, with high-fidelity physics and diverse, procedurally-generated environments. MolmoSpaces-Bench, in particular, is a test of controlled variation across a variety of pick, place, open and close tasks, plus compositions thereof.

It’s a new benchmark that’s definitely not close to saturated yet. And DreamZero performs well on both:

What Can We Learn?

Let’s compare pi-0.5 with DreamZero, specifically, as pi-0.5 is the second best.

Training data: pi 0.5 was trained on 10k+ hours of real robot data, VLM data, and data from the Droid dataset, as Mahi Shafiullah points out on X. DreamZero was trained on either DROID data or Agibot data, depending on the model checkpoint.

Training data distribution likely makes a huge difference here. Notice how in the DreamZero paper, Dreamzero radically outperforms pi-0.5 on AgiBot (not represented in the pi-0.5 training data) but is much closer on the shared DROID-Franka setup:

This also seems to suggest to me that possibly the additional 10k hours of robot data is not as useful as people think. Pretraining on the right robot data does seem very important; in another recent blog post, Physical Intelligence showed a massive gain from pretraining on in-distribution partner data:

So, maybe adding in an extra 10k hours from a different robot is actually no better than whatever random video data you happen to have. This would likely be a bad result for people who want to train cross-body robot brains. Perhaps you’re not getting anything from those other robot embodiments that you wouldn’t get from just adding cheap, plentiful egocentric video data.

Model backbone: DreamZero is based on Wan2.1-I2V-14B-480P, a 14B-parameter video generation model. It can take up to 8 frames of context. By contrast, pi-0.5 was trained based on the 3B-parameter PaliGemma open-source VLM. So, a much smaller model. It’s also one that takes only a single frame instead of an 8-frame history. This has ramifications for any task which is partially observable or has hard-to-model dynamics, which is to say any real robotics task.

Model size: DreamZero is a huge model, and much of the work in the paper is actually around making this massive, 14B parameter model run in real time. Ablations in the paper seem to show that model size makes a very large difference here:

The problem with both including more history, and with increasing model size, is that this would usually make the model both harder to train and cause it to “overfit” more in a low-data regime. Unlike LLMs, where overfitting really isn’t an issue due to abundance of data, in robotics we’re essentially always operating in a low-data environment, even now. DROID is tiny compared to even the smallest LLM datasets.

So, a proposal: the video-generation objective functions as an “auxiliary loss,” forcing structure on the DreamZero model and causing it to learn some internal model of the world. It provides a comparatively strong form of supervision compared to the sparse signal from robot actions. This in turn makes it better able to adapt to the varied MolmoSpaces environments that it was not trained on.

Final Thoughts

There is a lot we can’t determine just from these papers. We don’t have access to all the data that Physical Intelligence used; it’s still hard to access the GB200s that NVIDIA uses for inference. But to me, it’s exciting to see that perhaps we don’t need nearly as much data as was previously thought to get strong real-world robot results, and I hope to see more work in this direction.

One last question I have is this: if video generation is “just” an auxiliary loss, how long will we need it? In a world with thousands of intelligent robots deployed, will this eventually be distilled away, and will all of these models just be beaten by dedicated robot video models with robot-specific fine tuning? This seems very likely, but is potentially a couple years off.

Please let me know what you think below.

Subscribe now

Share

Leave a comment

Let me start with a vision of a potential future:

You wake up with a jolt in your foxhole, a little dug-out somewhere in the no-man’s land between two armies. A light is flashing on your beat-up tablet.

Maybe you hope for a bit that it’s a message from home, or even from command -- it’s been weeks since you’ve seen any other humans on your side of the war. Maybe there aren’t many of you left.

It’s not, though. It’s the enemy, moving through the trees a few dozen kilometers away, spotted by a little drone the size of your hand. At your command, a small army flies into action: the rotors of bomber-drones whir to life, the rolling autonomous gun-platforms taking up strategic positions in cover and to cut off any chance of escape. Robot dogs carrying guns sweep into nearby buildings, scuttling up bombed-out stairwells and setting up lines of fire. Anti-personnel drones loaded with explosives swarm through the trees.

Kill them all, you say, and your robotic troops spring into action. They know exactly what to do, coordinated by an artificial intelligence with superhuman ability to coordinate disparate data sources and make decisions.

You lie down and close your eyes. When you wake back up, the enemy are all dead.

In the battlefields of eastern Ukraine, drones are increasingly becoming autonomous. When given a target — a Russian soldier or vehicle — they will “lock on” and use their on-board artificial intelligence to pursue and kill that target. At the same time, defense startups have begun building combat-ready robots with increasing levels of autonomy.

In addition, the whole business of fighting a war seems liable for automation. The kill chain — the sequence of information-gathering, decision making, and action-taking that leads to deploying lethal force — is well suited for artificial intelligence.

Let’s talk about how and why.

Subscribe now

What is the Kill Chain?

In a war, soldiers do not cleanly agree to meet in a small area and shoot each other, as in a video game. Instead, enemies must be found; they must be identified; assets must be allocated to engage them; and then finally munitions must be used to destroy them. This is the “kill chain,” and it stretches from generals in Kyiv down to the men in the trenches on the front lines of the Ukraine war, or from US Navy admirals down to the sailors who fire missiles at boats in the Caribbean.

Large parts of this, especially in a modern military environment, look a lot like the sort of data analysis tasks for which AI is particularly suited. AI models like Anthropic’s Claude can analyze the massive amount of data produced by a modern military, identify trends, and respond faster and — one hopes — more accurately than human analysts can.

This has led the US military to use Claude for military operations, like the US strike on Venezuela that led to the capture of President Maduro. It’s also led to conflict this week between the Pentagon and Anthropic, as Anthropic has a “red line”: Claude cannot be used to make decisions to kill people on its own.

AI Agents for Combat

Deploying drone swarms will require autonomy. These simply cannot be remotely piloted in meaningful numbers, as right now, for the common FPV drones that inflict so many casualties in Ukraine, you need a single pilot per vehicle.

All of this will change. As autonomy gets better, it will be possible for one person to manage multiple drones, say in order to strike a position multiple times in rapid succession to pin them down or hit enemies attempting to reposition. One pilot will manage three drones, then ten, then a hundred.

And it won’t stop with drones. Some of these robots will be armed ground vehicles. We’ve seen recently companies like Tencore pitching their ground robots as part of combined forces which escort and work with humans on the ground, protecting them from aerial drones. Countering this threat will require superhuman reaction times, and likely push these systems to have more and more decision making authority on their own.

One of the clearest visions came from Scout AI co-founder and CEO Colby Adcock, who proudly announced the Fury Advanced Vehicle Orchestrator in an X post. This is an agentic system which can take a natural-language command and turn it into action. In their demo, they show a human commander giving a high-level command, and then AI agents orchestrating a response: finding, fixing, and destroying a target vehicle using a combination of ground and aerial robots.

To understand what’s different about Scout’s approach with Fury, you also need to understand that it’s using a fundamentally different method than military robots in the past, instead being based on a family of Vision-Language-Action foundation models.

These models can consume video feeds and sensory data from different robots, and aggregate them to build a single unified picture of a battlefield with which to make decisions. The commander says “find and destroy the red vehicle,” and the robots execute: they gather information, process it, and complete the mission.

Final Thoughts

For now, these robots are still semi-autonomous at most. The majority of combat robots in Ukraine are under full human control: they are remotely operated, their purpose is to keep a human soldier out of the line of fire while he does his job. The semi-autonomous drones in Ukraine right now are not terribly intelligent; they’re still not much smarter than a guided missile. The robot ground vehicles, as I’ve written about before, are largely holding trench lines and conducting ambushes. It will be a long journey before Fury is much more than a demo video for a VC pitch deck.

Besides, to an extent, autonomous weapons have always been with us. A land or sea mine is a crude version of an autonomous weapon, wired to blow when an enemy (or anyone else) makes contact with it.

Mines are also an example of the realistic failings of such systems: they make mistakes. A lot of mistakes. People will die for decades after a conflict ends due to mines or encounters with previously-unexploded munitions. Perhaps autonomous robots will be the same; perhaps when demolishing some derelict building ten years after the end of the Ukraine war, a robot sentry will activate and gun down a handful of workers.

There are other concerns, too. From Dario Amodei:

Of course, these weapons also have legitimate uses in the defense of democracy: they have been key to defending Ukraine and would likely be key to defending Taiwan. But they are a dangerous weapon to wield: we should worry about them in the hands of autocracies, but also worry that because they are so powerful, with so little accountability, there is a greatly increased risk of democratic governments turning them against their own people to seize power.

Autonomous military forces will require good governance, keeping them accountable and strictly under the control of civilian authorities. AI agents for military operations radically increase “AI risk.”

There are positives to these systems, too, of course — at least distant, potential positives. By taking human warfighters out of the direct line-of-fire, they could save lives on both sides: robotic systems are more disposable, after all. If a situation is ambiguous, it’s much more reasonable to lose a few $300,000 robot ground vehicles than a few of your human soldiers, so maybe you give the order to hold fire and wait for more information. These systems can also make decisions faster, using more information, and will perhaps one day be able to make those decisions better and with less error, matching the trends we have seen in other parts of society.

This is a uniquely strange and risky area for robotics, though. I write this mostly because I think it’s important everyone understand that this is here, and it is happening today.

Please like, share, and let me know what you think below.

Share

Leave a comment

Learn More

There was a saying popular in Silicon Valley around the mid 2010s: “The last driver’s license holder is already born.” There is even a blog using this as its name. After a lull, it seems finally clear that this statement is, more or less, true in 2026. Ten more years of this and no one will be driving their cars.

Right now, there are two big trends in self-driving as I see it:

• At the high end, Waymo (and its competitors) are offering a fully autonomous taxi service which scales quickly to new cities, is offering millions of rides a year, and is a full replacement for the taxi.

• Automakers like BYD and companies like Wayve are increasingly offering high-quality self-driving capabilities for consumer vehicles, in a similar vein to Tesla’s Full Self Driving.

Not only are cars getting more intelligent, but more of them don’t even need humans. And this extends beyond personal automobiles to semi trucks, with Aurora recently bragging they can make the 15 hour trip from Phoenix to Fort Worth autonomously. In China, companies like Neolix have begun deploying self-driving delivery vans on roads (to sometimes hilarious results).

In this blog post, I’ll go over some of the recent news on self driving vehciles. If you’re interested in this or other robotics content, please subscribe below.

Waymo Continues To Expand

Waymo has at many times in the past faced skepticism that it could scale quickly. But this concern looks increasingly dated; they’re currently deployed in 10 cities, either through Uber or their own app. They’ve planned expansions to over 20 additional cities, including cold-weather cities like Buffalo. To fuel all this growth, they recently raised a $16 billion funding round. The scale of the operation is unmatched outside China (Baidu is the leader globally). To couple with this expansion, they’ve announced their next-generation vehicle, the Waymo Ojai, built by Chinese automaker Geely, which should greatly bring down per-unit costs while offering a spacious interior.

Waymo’s Remote Operators

There’s been a lot of discussion lately about Waymo and its use of remote operators, especially after recent testimony to Congress. Waymo claims to use a surprisingly small number of operators (or “assistants” to use their preferred term): just 70 for a fleet of 3000+ autonomous vehicles.

These operators’ job is not to drive the vehicle. Instead, the operators will usually give fairly high-level instructions: telling a vehicle it can ignore an obstacle in the road, for example, or giving a sequence of waypoints for the vehicle to follow in order to get it out of a tight spot.

This system is not always perfect. A serious stress-test for Waymo happened during PG&E outages which resulted in traffic lights going dark:

Waymo vehicles stopped all over the city, blocking intersections and causing serious trouble for emergency vehicles. This was due to a spike in requests for help going back to their command centers from the autonomous vehicles, which seems to have taken down their system. They claim to have implemented new protocols to prevent this from happening again.

Likely, we will see growing pains like this continue. But remember this: we absolutely do know that Waymo is a lot safer than a human driver. The case for autonomy is clear.

Autonomy in Consumer Cars

Tesla’s autopilot continues to improve. But it’s not just Tesla: we see more and more consumer vehicles adding autonomy features — Rivian Driver+ and especially BYD God’s Eye are growing increasingly capable.

The case of BYD is particularly interesting. BYD is the world’s biggest automaker now, and its vehicles are consistently well-reviewed and well-liked. When it rolled out its God’s Eye 5.0 self-driving software, it did so across 2.3 million vehicles in China alone. God’s Eye uses a mixture of camera, ultrasonic, and radar sensors to achieve safe, reliable driver assistance.

According to BYD, this generates 160 million kilometers of driving data every day.

Legacy automakers are following along. Nissan, for example, is working with Wayve to deploy its self-driving software as ProPILOT on their new vehicles starting in 2027. Wayve has set ambitious goals, showing that its system is capable of driving in cities all over the world in 2025 — such as in their “90 cities in 90 days” initiative. This approach works without maps; they visited both to rural areas and to dense cities.

Wayve has a camera-only autonomy stack. What’s interesting about the Nissan approach is that it adds radar and lidar for “Ground Truth Perception,” adding extra safety guarantees on top of the camera-based autonomy from the Wayve Driver. With the low cost of lidar and radar, this would not be expected to add much to the cost of the vehicle; and the extra safety guarantees may be worth it.

Tesla Robotaxis in Austin

And of course, when speaking of consumer cars, we must proceed to discuss the elephant in the room: Tesla. Tesla has their own robotaxi service, and they recently rolled the first fully-autonomous, steering-wheel-free Cybercab off of their production line in Gigafactory Texas. They’re planning to expand their robotaxi program to seven new cities in the first half of 2026, up from just Austin and San Francisco.

However, Tesla has seen a spate of bad news lately. From AV Market Strategist:

The new crashes, from December 2025 and January 2026, include collisions with a fixed object at 17 mph, a bus while stationary, a heavy truck at 4 mph, and two backing incidents at 1-2 mph into a pole and a fixed object. Buried in the updated data is a revised report for a July 2025 crash that Tesla originally filed as “property damage only.” Five months later, Tesla quietly upgraded the injury severity to “Minor W/ Hospitalization.”

It’s not been all good news for Tesla in other areas either - for example, we saw a recent $243 million judgement against Tesla for an accident in Florida involving autopilot.

Tesla is taking a unique, extremely low cost approach, building a very cheap Cybercab that can be rapidly scaled. If the Waymo Ojai is in the $70k to $80k range, as I and others have estimated, the Cybercab is going to be probably a quarter of the price. This is a huge win for them, and is a result of a car with substantially less on board sensing and compute.

The big question is whether or not this is possible. Humans, after all, use quite a lot of compute to drive, at least; our brains are insanely powerful thinking machines. Vision-based autonomy is clearly possible, but Tesla might have trouble especially if they need to continue producing models for legacy hardware.

Still, the vision is incredible: I loved this video of a Tesla delivering itself to its new owner, for example; Tesla continues to sell the vision of the self-driving car better than anyone else.

Final Thoughts

Despite a few bumps in the metaphorical road, the progress towards ubiquitous self-driving cars seems possible. There are many forces driving this change: not just Waymo, Baidu, and the Tesla Cybercab, but the fact that the general level of autonomy on all cars seems to be going up year over year.

And I can’t wait. There are so many horrific stories like this one caused by people driving and making mistakes that cost lives. The world will be better off when this technology is everywhere. Again, these cars are a lot safer than a human driver. And with more and more cities seeing autonomous vehicles, I think a lot of the public opposition to these cars will fade away, defeated by the sheer convenience and the quality of the experience.

I hope this year we see honest, successful, high-quality competition for Waymo — from Nuro, Zoox, and Tesla — but this seems farther off, as all three companies seem significantly behind in the United States.

But my main prediction for 2026 is this: by the end of the year, self-driving cars will be boring. People’s fears will be largely assuaged, the recipe for scaling will be clear, and the technology will be broadly acknowledged as solved.

Share

Leave a comment

Learn More

The AV Market Strategist

Waymo's Shocking Data Release, Uber Invest $100M in Robotaxi Infrastructure

👋 Welcome back everyone…

Read more

5 months ago · 17 likes · 1 comment

We all have too many papers to read; in the past I have just dumped a bunch of related papers together with some light summary, to say why they’re interesting. This paper roundup is on evaluation and benchmarking, which has been a theme for me lately.

Benchmarking robotics is hard; collecting robotics datasets you can actually use is hard. This blog post is a short overview of a few datasets that might be of interest, both lesser-known ones that struck me as interesting and a few large and well-known ones for context.

You can also use this blog post to scroll through and see what the datasets people are talking about in a RoboPapers episode, actually look like.

If you like reading thoughts on robotics, please consider subscribing. Usually I write more in-depth summaries; this particular blog post is much more minimal and “stream of consciousness” than usual.

Subscribe now

Datasets

Open X Embodiment

An absolutely huge dataset of robotics data. At the time of initial publishing, it had more than 500 skills and 150,000 tasks included, with a wide variety of robot embodiments. It won the ICRA 2024 best paper, and has about half a million authors. Find it here.

DROID

An important subset of Open X Embodiment, and a dataset that’s still heavily used today. Partly because the hardware setup (above) was duplicated across so many universities. Contains a mix of office, kitchen, and home environments, with a Franka Panda arm and a zed camera on the wrist, plus a number of third-person views.

Check out the website for more.

Build: 10k Hours of Egocentric Data

The Build team has released 10,000 hours of video data. 192,900 monocular wide-angle egocentric video clips, collected by human workers in a variety of different environments.

You can find the data here on HuggingFace. And look at Eddy Xu’s X thread.

OpenEgo: 1107 Hours of Egocentric Data

600+ environments with 120 million frames. Task, language, and 3D hand pose annotations. Often these datasets do not have annotations, so this part is potentially really valuable.

Check out the original X thread from Ahad Jawaid. Github site for OpenEgo.

ShareRobot: Task Planning

A task planning dataset, with complex multi-step tasks and question answering. You can find it on HuggingFace here.

HUMOTO: Human Object Interactions

A large interesting dataset of human-object interactions. Not sure how useful it is, but given how much object interaction in whole-body robot control has accelerated recently, this seems worth a look. You can check it out here.

Benchmarks

Libero

You probably know this one; it’s one of the biggest.

With 130 tasks spread out across 4 task suites, it has a lot of variety, and has been used in a ton of research papers so far. The code is also open source.

Contains datasets with RGB information from wrist cameras, proprioception data, language, and PDDL scene descriptions here. These are all high-quality human teleop data, making this benchmark highly suited for learning-from-demonstration research.

CALVIN

A commonly-used benchmark by Oier Mees for language-conditioned task execution. It uses the Franka Panda robot and has a simple setup with some light variation in objects and coloration, together with a good variety in pick and place tasks. Find it on Github here.

RoboEval

A benchmark of 10 base tasks and 3000+ demonstrations which aims, somewhat uniquely, to provide unique, non-binary metrics of performance, in order to give an idea of how and why experiments are failing: there are trajectory smoothness metrics, it tracks environment collisions, etc.

Check it out here. Thread with thoughts; original X thread.

RoboArena

Real-world benchmarking via web socket to your policy. Grad students run evaluations in different university setups.

I’ve written about this one before, when writing about evaluations. I think it’s one of the best ideas for policy evaluation I’ve seen; though it’s still woefully limited in many ways.

See the X thread here, or check out the project site.

PolaRiS

A benchmark specifically aimed at sim-to-real correlation. Scan an environment with your phone, use Gaussian splatting to construct a scene, and use a set of tools to create a photorealistic simulation. Ships with a set of “real-to-sim” environments in which performance correlation with the real world was validated.

Check it out here. Robopapers episode.

REALM: Real-to-Sim Validated Benchmark

A benchmark specifically aiming to close the sim-to-real gap. Sedlacek et al. put together this benchmark designed to be a good simulation that closely represents the real world. See the post on this from Evangelos Kazakos on BlueSky.

Physics IQ

A video understanding benchmark - but of course video understanding is very closely related to robotics. Basically, how well can video models understand certain physical problems?

Check out the website and the thread on X. There was also an ICCV 25 challenge and associated workshop.

Conclusions

There are a lot of benchmarks and datasets out there; comparatively few (1) have gotten real uptake, and (2) show some actual correlation with real-world results. This is still an active area, and I personally am very excited about potential for real-world benchmarking like RoboArena (though it’s very expensive).

Ultimately the best benchmark is the real world — fortunately, with common platforms like the Unitree G1, we can expect people to open-source code that others can immediately use and deploy.

Subscribe now

Share

Leave a comment

Learn More

Sometimes, you need to raise a round for your robotics startup and the training run just didn’t go so well. Or maybe the CoRL deadline is a week away and the results just aren’t there. Never fear; you have options. Follow this guide and everything will work out just fine.

Let’s discuss how to make your model look as good as possible.

1. Don’t let other people run comparisons. Remember what happened with Llama 4: if other people can try your model out, they’ll quickly uncover its limitations. If you can keep your model secret, that’s best.

2. Control the environment of your demo carefully. Lighting, objects, initial robot configuration, and so on. This lets you overfit the demo scene and get really nice, smooth, high quality motions.

3. Never show your failures. This one might seem obvious — why would I show failures? — but if people can see where your model fails, they can start to see the limits of what you can do. Only strong robotics papers and results can show failures with confidence.

4. If you have to let other people run comparisons, choose the people carefully. Make sure they’re only happening in the right circumstances, ones your model supports as roughly in distribution. Absolutely don’t do what Physical Intelligence or NVIDIA do and open source your model so anyone can benchmark it.

5. When working on the results section of your research paper or blog post, you may be tempted to include some baselines. This is a good idea; just be careful to choose weak baselines so you look good. Octo is a great choice here; it was well publicized but had lots of limitations that weren’t widely discussed.

6. On the same note: cherry-pick your benchmarks. There are a ton of robotics benchmarks out there, and they all test subtly different things. Importantly, these differences are not obvious to people who are not familiar with the benchmarks involved.

Now, you may be thinking: “Chris, this is all great advice; but people will call me out if I follow these rules.” Don’t worry about that. There are so many little things which influence robotics performance: camera placement, arm configuration, object diversity, low-level controller implementation, and so on.

As long as you can make an argument your setup has to be slightly different from everyone else’s, you can get away with a lot. For example; a benchmark I like is RLBench, which turns out to be a very difficult one for many VLAs, and many successful methods on this benchmark instead use motion planning together with a higher-level goal prediction model.

Second, as a result of the above, robotics benchmarking standards are quite low relative to other areas of machine learning. Papers from very famous roboticists get away with these things all the time, and they’re under much more scrutiny than you are. You’ll be fine.

What to Look For In a Strong Robotics Release

1. Open source code and models. Not always possible, but always welcome.

2. Very diverse scenes and environments. Modern robotics learning methods are very, very good at overfitting to a small task distribution — a clean table, objects at most a few centimeters from where they started.

3. Don’t be afraid to show failures. We all know robotics methods fail a lot; showing these is a strong signal, and also helps qualify where it works and where it doesn’t.

4. Compare against the current best methods. Pick the best models head-to-head with embodiments and benchmarks they report on. Don’t cherry-pick.

On the same note, I think paper reviewers must be accepting of some results which aren’t as good as other methods; it must be allowable to fail at some benchmarks, or people will cherry-pick.

Some Caveats

Almost all of the robotics videos and results you see are, I believe, real — in that they’re doing exactly what the creators say they are. The problem is that because it’s so easy to overfit to a particular scene, and because the limitations of a model are so hard to ascertain from a 30 second clip, it’s really hard to tell whether a team is making progress toward the underlying goal of general-purpose embodied intelligence.

And robotics is hard; just because a team is employing some of the tricks I wrote here does not mean their results are invalid or their model is weak. I am certainly guilty of them all, at one time or another! One of the fundamental issues with robotics projects is that there are so many things that influence performance, that it’s very hard to distinguish the signal from the noise.

On the same note: lots of machine learning researchers from other fields don’t understand how hard robotics benchmarking is. They will often insist on building their own, usually simulated, benchmarks, which invariably don’t tell us anything and just add more options for #6.

In the end, robotics benchmarking will be solved by having lots and lots of robots, and models that actually work across most of them. More projects like Lingbot-VLA, DreamZero, and pi-0.5 — models that people can actually try out on different robots, use, and openly compare.

Share

Leave a comment

I wrote about evaluation a bit in the past, and will surely do so again:

The release of the Imagenet dataset was a landmark moment for the nascent field of deep learning. This collection of what is now 14,197,122 images covering more than 100,000 concepts (“synonym sets” or synsets) was key in driving and assessing early progress in deep learning, in part because it signaled the ongoing shift from “better algorithms” to “better data,” which in turn unlocked much stronger algorithms. Quintessential, landmark deep learning papers like Resnet and ViT were evaluated on ImageNet.

The dataset has many nice qualities: compared to other popular vision datasets like CIFAR-100, ImageNet has much more variety; it has high-resolution images; a method working on ImageNet tells you something (not everything, but something!) about whether it will work in the real world. But it’s still manageable: most “Imagenet” results stem from Imagenet-1k, a clean subset of 1,000 chosen object classes (“Golden Retriever”, “Sofa”, et cetera).

But this was a very, very controlled problem: image classification; i.e. “which of these 100,000 classes does this image that I am looking at belong to?” Image classification is easy: the problem is clean, it’s well defined, it’s not going to change or fluctuate. It is, in short, it fails to characterize systems which operate via repeated interaction with their environment as opposed to a one-off image capture.

And so we come to robotics. With the rise of humanoid robots and massive funding for real-world robotics research, it’s more important than ever to be able to tell what actually works and what does not — but at the same time, this is more obfuscated than ever.

Fundamentally, though, there are two main options: evaluate in the real world (somehow!) or evaluate in a simulation. Each have serious advantages and disadvantages. Let’s talk about them.

If you’re interested in seeing more of this kind of post, please like and subscribe.

Why an Offline Dataset Doesn’t Make Sense

First, though, let’s discuss the issues with using an offline dataset for evaluating robot trajectories. It works for images and language, after all — why shouldn’t it work here?

An offline dataset might take the form: Predict action given current state observation and task description. However, robotics is an inherently interactive domain; small errors in action prediction accumulate over time, and lead to different outcomes. Good policies compensate for these, and recover from partial task failures.

Without an interactive environment to evaluate in, we can’t compute task success rates, and we can’t determine whether a policy would be useful at deployment time. This leaves us with two options: (1) test in an interactive simulation, and (2) find a way to compare methods on real hardware.

But Simulations are Getting Better

If we need interactivity, perhaps the best benchmark, then, will be in simulation. And simulations are getting more powerful, more interesting, and more diverse all the time. More importantly, we’re regularly seeing previously-impossible examples of sim-to-real transfer. See, for example, Doorman, by NVIDIA, which was sufficient to teach a robot to open a door — though note many difficulties involved in this work!

Simulations are getting more powerful and easier to use all the time. But few of these simulations rise to the level of a usable benchmark, i.e. something like Chatbot Arena, Humanity’s Last Exam, or SWE-Bench Verified.

For robotics manipulation, two of the most notable benchmarks are Libero and Calvin. These implement a wide range of tasks with language conditioning (“push the button", “stack the blocks”), which means they can be used to train and evaluate multi-task policies. For mobile robotics tasks, the most notable simulation benchmark is Behavior 1k, which implements a thousand challenging simulated household tasks.

But there are many more, and they all have subtleties which impact which methods work. This makes it easy to “choose your own” subset of benchmarks upon which your model will perform the best, which renders moot the whole point of even having benchmarks in the first place! Efforts have been made to unify all of these different simulators, like Roboverse.

Another issue is that authoring tasks in simulation is hard. In the real world, if I want to have the robot stack blocks, I go buy some blocks and drop them in front of the robot. In a simulator, I need to get the friction parameters right, masses, make sure contact is working properly, et cetera. I need to implement cost functions (for RL) and success criteria, and this only gets harder as I start to scale simulation up. One notable effort to reduce this pain point is the Genesis simulator, which attempts to use user-prompted natural language to help create environments.

Fundamentally, though, simulations are still hard to work with and an inaccurate representation of the true robotics problem, without sensor and actuator noise, with overly-clean problems, and with unpredictable, often-inaccurate contact dynamics. As a result, there will always be a role for real-world evaluation.

Evaluating In the Real World

Obviously, comparing performance on a real-world task for robots is the benchmark. But running any kind of evaluation in the real world for robots is hard. In simulation, you don’t need to “reset the environment,” putting everything back where it was before you run again — something I have done many times in my life as a robotics researcher. Fortunately, we’ve seen a couple ways in which this problem might be addressed, especially inspired by recent work in large language models.

When the AI field moved towards large language models, we quickly saw a rapid proliferation in the number of benchmarks of ever-increasing difficulty - benchmarks like Humanity’s Last Exam fit the Imagenet mold of: here is a dataset, see if you can get the right answer.

But benchmarks quickly saturate; and they never solve the “real” repeated-interaction, high-dimensional data problem we actually care about, whether the goal is language or a robot. One evaluation method which exploded as a result was Chatbot Arena: a platform in which a user comes up with their own prompt, it is sent to two different LLMs, and the user chooses whichever response was better. While the particular implementation has not been without issues or critics (especially notable is Llama 4’s apparent benchmark-maxxing), the approach is scalable, in that it doesn’t require running a full sweep of all possible queries every time. While it’s not perfect, because no two evaluations are the same, it gives you an ELO rating which gives an idea how competitive every model is with other options out there.

This is a great fit for robotics, where similarly, running individual evaluations tend to be extremely expensive. Tournament-style evaluation most useful because it minimizes the number of expensive evaluations you need to run in order to see if it works; crowdsourcing queries also helps prevent overfitting to the benchmark (which was a perennial problem in a lot of computer vision research, and persists in many fields to this day).

Examples include RoboArena, which is a community-run cloud service for evaluating robot policies. Your policy gets executed on the cloud, and you just need to provide a service exposing it. This does limit the kinds of tasks that can be evaluated, though: latency will always be a serious issue.

The authors of the large humanoid robot dataset Humanoid Everyday are also planning a cloud service for evaluating robot policies on a real Unitree G1 humanoid; you can check their site for details (as of writing, it still says it’s coming soon) and watch our RoboPapers episode on Humanoid Everyday to learn more. You can also watch a podcast episode on RoboArena while you’re at it!

Common Real-World Platforms

All this is very limiting, though. Maybe we should just expect that people will just own robots upon which standard policies are expected to work, so that they can do their own evaluations on their own problems.

It might even be possible for something open source to take over, something that I’ve written about in the past. The HuggingFace SO-100 arms have seen some significant uptake in the hobbyist community, though very few have made their way into academic research papers that I’ve seen. We might see mobile versions of such a platform succeed; XLeRobot, for example allows for you to, at a fairly low cost, test out different mobile manipulation policies.

Of particular note today is the Galaxea R1. This platform is available for a relatively low cost, and comes in several varieties to support easy data collection, like the R1 Lite. The R1 Pro was used in the Stanford Behavior Challenge at Neurips 2025. You can download a VLA and associated pre-training data for it. It was even used in the recent RobbyAnt LingVLA, which used 20,000 hours of real-robot data.

The most common platforms right now are probably the Trossen ALOHA arms, the YAM arms from I2RT, and of course the Unitree G1. It’s notable how compared to previous systems — available even 2-3 years ago — all of these robots are spectacularly cheap. Robotics has become substantially more affordable, and robotics research more accessible. As a result, maybe the best way of telling which methods are “good” is just to watch to see which methods people build off of when performing experiments with these common platforms.

Final Thoughts

So let’s recap:

• Offline datasets do not work because robots never do exactly the same thing, these errors compound, and all robot tasks are too multimodal for this to be meaningful

• Simulations exist and are useful, but are niche, hard to implement, and often missing critical aspects of the real world (usually visual diversity, implementation of interesting/relevant tasks, and high-quality contact simulation)

• Real-world evaluation is slow and horribly expensive to run, and can’t match the expectations of other AI fields like language or image understanding in terms of speed.

This sounds dire, and it actually gets somewhat worse; because all of this has focused on assessing algorithms, and robots are not merely algorithms: they’re hardware plus an algorithm. Hardware factors — joint positions, sensor placement, motor types, backlash, heat buildup, and more — often matter more to task execution than the algorithm you’re using.

So any “real” benchmark comparison of, say, Figure 03 vs. Tesla Optimus vs. 1x NEO would by necessity look more like a human competition, where participants, say, go to a test kitchen and see who can load a dishwasher the fastest.

We’ve seen the early evidence of such events, like the World Humanoid Games, or the many competitions we see every year at major robotics conferences like IROS and ICRA. These are likely to expand, although major companies right now have too much to lose and too little to gain to bother competing.

On the model side, the fact that you can just download and run pi-0.5 on your hardware, for example, is an incredibly promising start. In the end, though, the answer to the question, “how do we quantify progress in robotics?” has to be “all of the above.”

Share

Leave a comment

Compared to other humanoid robots like Figure 03 or 1x NEO, Atlas is an alien. The product version of the storied humanoid robot from Boston Dynamics has strange-looking, bowed legs; it has an odd, circular head like a lamp, and all its joints can rotate all the way around.

If you want an illustration of just how strange this looks, watch this video from user CIX on X, recorded at CES 2026 in Las Vegas:

Contrast this with a humanoid like the Figure 03, which was clearly designed to mimic the appearance and capabilities of a biological human, something that I’ve covered before in a previous blog post.

Both of these robots are incredible pieces of hardware, but we must ask, why should Figure’s robot look so human while Boston Dynamics opts for such a strange form factor? Is it just a gimmick that Atlas can turn all the way around? If we’re moving away from the human form, why not just go all the way and make a robot that’s fully optimized for its task like the Dexterity Mech (video from Dexterity):

They do, on occasion, call this beast an “industrial superhumanoid,” and it’s a dedicated pick-and-place monster with a 60kg payload.

So let’s talk about why some of these robots look more or less human than others, and what the pluses and minuses are, with a particular focus on the new design from Boston Dynamics.

First, Why A Humanoid?

When asking why use a humanoid at all, the real question you’re asking is usually “why legs?” And this is an important question; lots of the robots, including the Dexterity Mech shown above, do not need legs. What are legs for, then?

Well, legs allow robots to:

• Handle more complex and challenging terrain

• Carry heavy things without requiring a large base

The first benefit is obvious — legs allow your robot to climb stairs or cross a debris-strewn landscape. They mean that your robot can be deployed in a wide variety of environments with far less concern about “preparing” the environment for robots.

This, however, rarely going to be a deal-breaker for real-world deployments, as it’s already economical to design industrial spaces to optimize productivity. Amazon, for instance, famously developed new techniques for creating flatter floors in its warehouses to benefit its Drive robots. So, in a real, large-scale deployment, legs are of use for handling terrain — but that’s only a very limited use.

More important is that legs allow robots to be smaller while performing the same tasks. Or, more accurately, it’s because a bipedal robot can perform the same work in a smaller, more constrained area. Because they’re dynamically stable and omnidirectional, an industrial humanoid robot like Atlas can carry a heavy load with a much smaller footprint than a robot like the Dexterity Mech.

This is important because in a factory or warehouse, you’re often trying to fit as much stuff into available space as possible — you don’t necessarily want all the extra space it needs to make a large, high-payload wheeled robot work.

But notably, this does not mean that your humanoid has to look human.

The New Atlas

The new Atlas robot has a couple unique design features, courtesy of product lead Mario Bollini on X:

And the unique legs:

With only two unique actuators, the supply chain and cost of the robot can be greatly reduced versus a more human-like design. The fact that the legs can bend forwards or backwards gives it some more flexibility, but also means that the legs are swappable left to right.

And with that in mind, let’s go back to that first video by CIX, where the (prototype, not mass production) Atlas reverses itself during a procedure. Remember when I said the main advantage of a humanoid was working in a more constrained space? This design grants the robot distinct advantages in constrained environments.

Why To Be Human

Contrast this with the approach taken by competing American humanoid manufacturers like Figure, 1X, or Tesla. Their robots are very closely designed to match the human form factor.

There are a few advantages to this:

• Teleoperation is easier. Even if your robot is superhuman, the humans operating it are not — and teleoperation is already pretty hard work!

• We have lots of human data already. The internet is filled with human video data; training from this data, as 1X has done, allows you to easily resolve the robot data gap.

• Our tools and technologies are all designed to be used by humans. This is a favorite argument of Elon Musk, for example. If your robot is expected to use tools or drive a forklift, you might want it to look human.

• It looks and acts human, and people like that. Robots that work around people need to be pleasant and likeable; people might not want to purchase these strange, scary alien beings whose heads can rotate 360 degrees.

There’s a safety angle to this final point as well; if a robot’s capabilities are roughly human, people know what to expect from it, and it’s important humans have a good model of what robots can do if they’re going to be working and living alongside them. This is a huge part of the justification for the design of the 1X NEO, which has very humanlike lifting strength and capabilities.

Personally, I find that there are a lot of holes in these arguments.

Robots Using Tools

I don’t buy that, in the future, we’ll want humanoid robots to use tools built for humans. When transportation in human cities switched from being dominated by horses to dominated by cars, every piece of our infrastructure changed. This will happen with robotics, too, as robots supplant human labor.

It seems very unlikely to me, for example, that humans will be buying non-robotic forklifts for their warehouses in 10 years. Every forklift will be something like a Third Wave robot; you certainly won’t be asking Optimus to go and drive a forklift, because the extra sensors necessary for automation will be extremely cheap.

The same will go for tools; maybe robots will have swappable end effectors, or maybe tools will be specifically designed with attachments for robot hands, but there’s no reason not to think that, at scale, you gain more from good, vertically-integrated design than from building something to support legacy hardware (humans) forever. Indeed, a modular robot like Atlas could eventually use these tools better than a human ever could.

At best, I think the robot tool use argument will be a short-term cost-saver that applies over the next couple years.

Is Data Collection Better With Humans?

This is a vastly better argument in favor of human mimicry, but the cracks are starting to show even here. Human teleoperation data, while essential for robot learning to this point, will not be able to take full advantage of superhuman humanoids.

But there are ways around this, and one is something we absolutely need no matter what: reinforcement learning. Says Atlas product lead Mario Bollini again:

Reinforcement learning is crucial for real-world reliability, as demonstrated in recent works like Probe-Learn-Distill from NVIDIA and RL-100 (both are RoboPapers podcast episodes you can watch/listen to). It also provides a way for us to start with human demonstrations but then improve upon them.

But what about human video data? Certainly, there’s compelling evidence that video data can improve performance with humanoid robots. I’ve discussed the importance of co-training on this blog before: how else could we ever we get enough data to train a Robot GPT?

But your robot might not need to be human to take advantage of video data. Take a look at the the recent “Emergence of Human to Robot Transfer in VLAs,” by the Physical Intelligence team. In the plot above, the show how just by co-training on human and robot data, the models naturally learn a similar embedding space for tasks which is shared across the different embodiments.

And PI’s robots do not look remotely human! They’re very simple, lightweight research arms with two-finger grippers. Now, they’re not performing highly dexterous tasks in this case, and this might change, but I see no reason why as long as the robot hardware is capable of a task, that such a shared mapping cannot be learned.

As a final note: I don’t think there’s any particular reason human hands have five fingers. Dogs have five fingerbones, as do whales; neither of these animals use these at all. Humans have hands with five fingers by accident of evolution, nothing more — it is not the product of some optimal engineering process. And so I don’t see why our robots should be limited to that, either.

Final Thoughts

The humanoid form factor is, I think, here to stay due to its clear advantages, but that doesn’t necessarily mean that it will stay human. The Atlas is an interesting look at a very different vision of what a humanoid robot can and should be, and I think it’s exciting to see it come to fruition with a new model designed for mass production.

I also think there’s a huge opportunity here. As I mentioned above, one advantage of robots that look human is that you understand what humans can do. With a totally “alien” design like the new Atlas, the roboticists can rewrite the script: you know what a human can do, but also what Atlas can do. That kind of product identity will, I think, be very valuable as we approach a sort of “jagged” physical AGI in the coming decade or decades.

Please let me know your thoughts below, and share/like/subscribe to help others find this if you found it interesting.

Share

Leave a comment

Futher Reading

On December 23, 2025, the Ukrainian military announced that a robot from DevDroid, a domestic producer of combat robots, had held a position near Kharkiv for roughly 45 days against sporadic Russian attacks. Employed by the 3rd Separate Assault Brigade of the Ukrainian Armed Forces, this deployment is emblematic of how these robots are becoming more important in the grueling attritional warfare in Ukraine.

This particular event marks something of a milestone for ground robots, which are used by Ukraine for roles like reconnaissance, resupply, infantry support, and rescuing wounded personnel. Ukraine’s military is particularly focused on robotics deployment, with their goal being to deploy 15,000 unmanned ground vehicles (UGVs) — among them at the very least hundreds of DevDroid platforms — by the end of 2025.

This might seem strange to some: after all, aerial drones are the iconic face of robotic warfare. The DJI Mavic is particularly iconic; as one Russian military blogger recorded, “Mavic means death.” This sentiment — and the overwhelming superiority of Chinese dronemaker DJI — led the United States to implement a ban on the import of such systems from China. Ukraine’s Operation Spiderweb and Israel’s decapitating strikes on Iran during Operation Rising Lion would not have been possible without these transformative platforms.

But there are roles for which these smaller kamikaze platforms are not particularly suited: the war is currently a grinding attritional battle, with brutal trench fighting that has consumed hundreds of thousands of human lives already, leaving both sides scrambling for more personnel to fill the ranks. And, of course, reaching for technological solutions.

Subscribe now

Why Ground Robots?

We can go on and on about the potential of small, disposable aerial drones in warfare (read my previous blog post, for example). These lightweight, low-cost machines are effective at eliminating enemy infantry and armor; they provide reconnaissance support; they mount ambushes. The Armed Forces of Ukraine received about three million first-person view (FPV) drones in 2025.

But these are not the only aims for warfare. In the end, war is about holding ground. Aerial drones may excel at engaging enemy forces, but arguably they’re just acting as a lightweight replacement for artillery. They can certainly engage the enemy; it’s been reported that drones cause 70% of casualties in Ukraine.

Note, however, that in World War II artillery and air strikes caused 50-70% of casualties. It’s not like the drone is replacing an infantryman. Inflicting casualties, largely, is not their job and has not been their job through much of history. Instead, as Napoleon Bonaparte said:

The hardest thing of all is to hold the ground you have taken.

This is where I think these ground drones come in, and why they have a very distinct role compared to the aerial variety (loitering munitions). Much of the Ukraine war is actually what appears to be almost old-fashioned trench warfare, with human soldiers digging in to hold territory against their enemies.

What These Robots Are Doing

It may surprise some users that we even want or need ground robots instead of just relying on swarms of drones, so let’s go into what these robots are actually doing a little bit more. There are two main classes of armed unmanned ground robots from DevDroid:

• The machine-gun-armed TW 12.7, recently approved for use by the Ukrainian Ministry of Defense

• The NW 40, armed with a rapid-fire grenade launcher, used for ambushes against light armored vehicles and enemy convoys, and recently codified (officially inducted into service for state procurement)

In addition, we see many other types of ground robots — kamikaze ground robots, logistics/supply carriers, and others for evacuating the wounded. Many of these are useful for mounting ambushes or for protection against them (in particular using autonomous robots to resupply).

Probably the most important role these robots are serving is to keep soldiers out of line of fire. Robots like the Devdroid or the similar T-700 carry fairly heavy weapons. They can perform fire missions and suppressive fire, and conduct ambushes without exposing human soldiers the enemy. And a lot of what human infantry do, unfortunately, is dig in somewhere and get shot at.

When a robot is "holding a trench” for 45 days, it won’t have spent all of that time taking enemy fire — but it was important to that someone be there, at that intersection of trench lines, who could fire on approaching enemies and take fire in turn.

This process of trading fire creates friction between the two opposing forces, and most likely once the robot starts shooting, the enemy would fall back. If they actually wanted to take that position they might call for artillery or — yes— aerial drones. If no one was there to oppose them, they might just break through, and be able to interfere with supply lines or just seize ground, dig in, and call for reinforcements.

Both sides will start shooting well before any casualties are guaranteed or even likely, at least in any particular engagement. In the American War on Terror, US forces expended something like 250,000 rounds of ammunition for every insurgent killed. But the act of firing on a position — of putting vast amounts of lethal power downrange — forces soldiers into cover or retreat to avoid harm. It prevents movement, and locks down a whole area, even if the engagement between the two forces was very brief.

And this is why a bomb-armed quadrotor — or even 50 quadrotors — can’t do the same job. The gun-armed robots we see here can move around, take some measure of cover themselves, and are armored enough to be quite resistant to explosives and even the occasional suicide drone. They can threaten a very large area — engagement ranges for these weapons run from 50 to 800 meters — and prevent enemy movement through a large part of that area, for days or weeks at a time, all without endangering soldiers on their own side.

It’s a very different part of the puzzle that is a modern battlefield: as opposed to being disposable, high-precision “light artillery” like aerial drones, these robots are acting like infantry or “light tanks.”

Autonomy on the Horizon

As noted above, these robots are not autonomous. But this is likely to change, and it might not be very long before it does — and the reason comes back to aerial drones. Basically, today’s militaries badly need an economical solution to keeping the “lower skies” clear and their troops safe (or, at least, as safe as it gets in a warzone).

Currently, counter-drone efforts are largely a manual affair, using specially-equipped interceptors which, themselves, are human-piloted FPV drones. More autonomous and scalable solutions exist, but often interceptor missiles end up being more expensive than the robots they are shooting down!

This has led companies like Anduril and Allen Control Systems to start building autonomous gun platforms which can automatically detect, target, and shoot down fast moving drones. Think, similarly, to the Phalanx CIWS on American carriers: these shoot a lot of bullets (relatively cheap) to take out the incoming drone while protecting humans on their team.

There is, rightly, a lot of debate about robots making the decision to shoot against humans, and the United States Department of Defense still says humans will stay in the “kill chain” at all times. But if they’re just shooting down munitions, presumably, there is no issue.

Of course, as countermeasures (to remote operation of these robots) grow stronger, and artificial intelligence gets better, who’s to say that robots won’t start to be trusted with more and more autonomy when handed other targets?

Final Thoughts

The old cliche goes that necessity is the mother of invention. Many of these current combat robots exist to solve for very specific problems the Ukrainian military faces, particularly a lack of manpower.

But these still feel like they presage future warfare in important ways. These machines will appear first to help hold ground, and to secure the lower skies against enemy drones — but if the need arises, they’ll be doing more than that.

If you like this, or have any thoughts, please let me know below.

Leave a comment

Subscribe now

Share

To learn more, read my previous article about aerial military drones below.

When a technology finally clicks, the changes spread faster than anyone expects and have far more serious implications for society. I want to lay out a case here for robotics finally “clicking” within the next few years, and what that could look like for everyone. Technological change happens fast, and nowhere is that more obvious than in the early part of the 20th century.

There were 130,000 horses in New York City around 1900. By 1912, they were already outnumbered by automobiles; today, there are only 68 licensed carriages and probably a mere 200 horses in the entire city now. Despite peaking at 26 million, the United States population of horses went down to 3 million by 1960. As soon as the economic utility of the animals vanished, they disappeared from public life and the city changed forever essentially overnight.

In 1908, maybe 1% of American households had a car. This number tripled in the 1920s, going from 8 million to around 23 million. By 1948, half of all households had a car; by 1960, it was 75%. The shape of American cities changed completely over this window.

Predicting the future is very difficult, and the core problems in robotics are far from solved, but I think we’re seeing a very similar period of rapid change happening right now. The level of investment and growth in robotics and AI is reaching a fever pitch, well beyond what I expected 1-2 years ago. And, perhaps more importantly, most of what I have believed are substantial blockers to robotics deployments now seem solvable on a technical level:

• Robots are relatively cheap, mass-produced, and of increasingly high quality, with a robust supply chain.

• Data issues that have stymied robotics learning in the past look addressable.

• Core learning technologies — both supervised training on large datasets and reinforcement learning in the real world — have been proven out.

All this means that we should see robots in a lot more industries and parts of society than we’ve ever seen in the past, so let’s talk about the future. But first, let’s lay out the case for robotics optimism — then we can get into what it means.

Subscribe now

The big story of 2025, for me, is the sheer scale of production of humanoid robots. Companies like Agibot and UBTech are building humanoid robots by the thousands now, and sending them to work in factories belonging to the world’s biggest automakers — companies like BYD.

In general, the number of humanoid robotics companies, and teams working on humanoid robots, is skyrocketing. Most recently, Rivian announced that it is spinning off Mind Robotics, with $115 million in funding. Said founder and CEO RJ Scaringe:

As much as we’ve seen AI shift how we operate and run our businesses through the wide-ranging applications for LLMs, the potential for AI to really shift how we think about operating in the physical world is, in some ways, unimaginably large.

An explosion of investment like this is never due to just one factor. In fact, several things have come together to produce this moment. Quality robots are getting incredibly cheap. The Chinese supply chain is getting very strong, making it easier for new entrants to build at least a v1 of their products. Hardware expertise is getting more widespread.

Techniques for robotics control and learning have become more mature, and have overcome a few major limitations that we’d seen in the past around mobile manipulation and reliable real-world performance. Partly as a result, companies like Unitree and 1x have demonstrated that there is real demand for robots from consumers, with preorders and widespread hype for 1x and exploding G1 humanoid robot sales for Unitree.

Finally, it seems that the robotics “data wall” is becoming less of an issue. Data collection and scaling is much easier than ever. A number of companies like Build and MicroAGI have appeared to scale up human-centric data collection; research work like EgoMimic has provided at least a feasible route to collecting data at scale (watch our RoboPapers podcast episode on EMMA here). Companies like Sunday Robotics are demonstrating how effective scaling with UMI-style tools can be (see our DexUMI episode on RoboPapers here).

Good Robots are Really Cheap Now

When I was in graduate school, this thing cost like $35,000:

And I’m talking just the robot arm - the Universal Robots UR5 - not the gripper, camera, monitor, GPU, et cetera. It was a pretty good platform, but back then the Robotiq 2-finger gripper alone was also probably about $12,000 — totalling about $62,000 when adjusted for the inflation we’ve seen since 2016.

These days, I could buy four Unitree G1s for that price (base model), or, for a slightly fairer comparison, one Unitree G1 EDU plus. I could also buy a LimX Oli. I will soon be able to buy a Unitree H2, a full-sized humanoid robot (probably $60,000-$70,000 USD, tariffs included).

Buying incredibly powerful, capable robots has never been easier: for less than the price of a mid-range sedan, you now can purchase a robot that was impossible even with DARPA funding just a decade ago. And all of these robots have a much more robust ecosystem: the robotics community has, practically overnight, become absolutely massive. Mass production has given us robots which are both cheaper and substantially more capable — as well as just more fun — than was true less than a decade ago.

And this has downstream effects, because a huge part of what makes it hard to automate stuff just comes down to price! Specialized hardware is expensive, but you need specialized hardware less and less — there are so many more options now. The expertise is more available. The hardware is more robust and easier to use. Lower cost makes the entire robotics ecosystem much stronger than it used to be.

We Now Have a Roadmap to Useful Robotic Manipulation

We’ve all seen a number of incredible dancing and martial arts videos from companies like Unitree. While these are impressive looking, they don’t demonstrate fundamentally useful robot capabilities, because they don’t interact with the environment. It’s relatively easy to program a robot that doesn’t need to collide with stuff; any dumb technique you can think of these days will work for building a robot that doesn’t need to interact with the environment.

Interacting with the environment, though, is tough. It’s hard to simulate; it requires a ton of real-world data to adapt properly. Getting the correct examples of physical interactions is very hard. I’ve spent much of my robotics career working on long-horizon robot tasks, and at this point the problems are very rarely due to high-level planning (the robot decided to do the wrong thing), but more often due to difficulties with environmental interaction.

You can take a look at the plot above from the OK-Robot (a 2024 paper) to get a sense for what I’m saying. A lot of the time, the robot can’t reach something (navigation failure); other times the hardware fails; other times it just can’t reliably grasp something due to a model not properly handling the particular combination of environment and object.

If robots could be made to reliably perform real-world manipulation tasks beyond structured picking, this would be a big deal. So I want to draw attention to two specific results that were extremely important this year: whole-body control with mobile manipulation, and real world reinforcement learning.

The video above is from RL-100, recent robotics research from Kun Lei et al. from a variety of institutions but mostly the Shanghai Qizhu Institute. It shows a robot arm working in a shopping mall — an out of distribution environment — while juicing oranges for seven hours.

We’ve similarly seen work like pi-0.6* from Physical Intelligence, which showed robots performing tasks like building cardboard boxes and making espresso drinks, reliably, in a way that humans might. And I’ve already written about folding clothes — startup Dyna Robotics started there but has since moved on to demonstrating high reliability in end-to-end learning with other tasks. Now, none of these tasks are revolutionary on their own, but achieving these levels of reliability with end-to-end systems in the real world absolutely is.

More importantly, the idea that there will soon be a recipe for deploying such skills is crucial. By analogy, look at ChatGPT and the broader adoption of LLMs; previously there were many different image- and text-recognition tools. Model large vision-language models like ChatGPT removed the barrier to building systems that leverage text and images; they provide sort of a shared interface. Even if that means running some on-robot RL procedure where someone types 0s and 1s for failures and successes into a spreadsheet, it seems that we could arrive at similarly highly useful systems for robots.

The Data Wall is Falling

But these advances, just like the ones we’ve seen for large language models, are all likely to be predicated on strong base models, just as we have seen for LLMs. Famously, the reinforcement learning that made Deepseek R1 such a breakout success was only possible because of the quality of pretraining making the reinforcement learning problems tractable.

The base models of robotics are usually called Vision-Language-Action models or sometimes Large Behavior Models (there’s actually a small difference here, but they’re accomplishing the same thing). I’ve also written about it as a direction for future VLA research. The question has always been where to get the data.

What’s changed is that now broad, diverse robotics data seems achievable, through a collection of different tools. While certain things seemed to be false starts (mass teleoperation data is too expensive, mass simulation often too difficult to tune for contact-rich tasks), there are other really good options which have come into their own this year: egocentric video data and “UMI”-style tool data.

Generalist goes into detail on their recent GEN-0 blog post, describing how diverse pretraining data leads to faster posttraining of their robot models. We’ve also seen from Physical Intelligence that as models scale, they start to learn to treat human data as “just another modality,” meaning that they can start to leverage it to improve performance on various skills. At some point, once we have enough data, this implies that a lot of data-related problems may fall away far faster than previously expected.

A number of companies have sprung up around this idea, including Build, which produces huge quantities of camera-based manufacturing assembly data from human workers, and MicroAGI, which gets rich, high quality data from workers in different industries.

Broadly, it seems more likely than ever that the “eternal” robotics problem of not having enough data, will be solved in the coming years. And unlike in my previous estimates, I no longer believe it will necessarily be some billion-dollar project — which means many companies will be able to build powerful autonomous systems.

The Elephant in the Room: LLMs

Large language models have also made dramatic progress in the last year. OpenAI’s o1 — the world’s first “reasoning model” - launched towards the end of 2024. Its success led directly to the release of Deepseek R1, which was a spectacularly important paper that described publicly a lot of the “secret knowledge” kept in house at OpenAI and has allowed for waves of successive exploration.

And the changes have been extreme. Coding is incredibly different from what it was just a year ago; it will likely never be the same again. Never again will I just write a whole project, token by token, by hand. These changes are appearing in many different industries: people raise concerns about AI replacing lawyers — increasingly many lawyers just draft everything with ChatGPT anyway. Similarly, 67% of doctors use ChatGPT daily, and 84% of them say it makes them better doctors. GPT 5.2 was evaluated on GDPval, a set of economically-valuable tasks, where it achieved equal or better performance than an in-domain human expert 70.9%.

Large language models are already significantly changing the way people work and rewriting the economy, like it or not. And these changes seem to be propagating far faster than the changes we saw with automobiles at the beginning of the 20th century, having propagated through society with all the speed of the internet. Robotics won’t spread so fast, but unless serious obstacles manifest (such as a total collapse of funding for R&D), it seems plausible there will be similar changes in the physical world.

The Next Few Years in Robotics

To go back to the metaphor at the start of this blog post: when horses vanished from American cities, they didn’t just lose their jobs. The whole infrastructure of cities changed: hay and stables were replaced by gas stations and parking garages. Manure was gone from the streets, replaced by exhaust fumes. The feel of a city, even walking around on foot, now revolves around cars, with wide, flat asphalt roads and traffic lights on every block. Similarly, if the “optimistic case” from this blog post holds true, we should see significant changes in the fundamental details of life.

So, to recap: we have seen a year of dramatic robotics progress which for the first time showed the viability of long-running end-to-end robotics manipulation, at the same time the cost of robotics hardware is collapsing and its quality is exploding. We have seen dramatic changes in the world of purely-informational artificial intelligence, through reasoning models and agents.

I believe I have motivated this “optimistic case” well enough now that I am allowed to speculate a bit, as I promised at the beginning of this blog post. To start, we may make a couple assumptions about robotics over the next several years:

• The robotics “data gap” will continue to close, and robotics will start to pick up speed due to the combination of more robots and more tools for making robots possible to use and deploy

• AI will be at the heart of this — both reinforcement learning and imitation learning will be key parts of the solution, as elaborated in my previous blog post on VLA research directions

This means that a lot of areas of robotics which were previously inaccessible to automation soon will be. In fields like construction, we have thus far been limited to incredibly simple and structured pieces of automation: specially-built roofing robots, for example. Similarly, much of manufacturing is already automated, but that automation relies heavily on specialized systems, sensors, end-effectors, and machinery. All of this makes automation extremely expensive.

At the same time, the labor markets for fields like construction and manufacturing are getting worse. The world has filled up in a way; our countries are graying. We expect the ratio of workers to retirees to go down, requiring each individual worker to be far more economically productive. Those same retirees will need care and companionship that humans are unlikely to be willing or able to provide. All of this means that the world of 2030 will be far more robotic than that of today.

I expect we will see far more robots essentially everywhere. Waymo and its competitors will expand to more and more cities; fewer people will use their personal vehicles to get around. If they do, their vehicles will be using Tesla or Wayve autopilot systems to get around.

Robots will be in homes. They may or may not be humanoids. But they’ll be able to perform a wide range of simple manipulation tasks, things like picking up and perhaps putting away the dishes. They’ll cost less than a car, possibly in the $10,000-$20,000 range for a very good one. Robot production will still be ramping up this point, so it’s probably still less than 1% of households that have an in-home robot — but it’s going to be rapidly increasing into the 2030s, until it reaches similar levels to cars by 2040, with 50% or more owning an in-home robot that can help do chores.

These home robots will often be companions first; modern AI is extremely good at companionship. It is striking to me how much for example my two year old daughter likes to interact with even very simple home robots like Matic and Astro; more capable and intelligent LLM-powered home robots will be far more compelling friends and “pets.”

Most importantly, this will also lead to the “iPhone moment” for robotics, which is when acceleration will really take off. Some — such as Sunday Robotics founder Tony Zhao — have publicly alluded to this moment coming. What we hope is that by 2030, users will be able to create and share their own use cases, as this is where true explosive growth starts to happen.

In manufacturing and industry more broadly, we’re already seeing “systems integrators” start warming up to widespread use of end-to-end artificial intelligence. As the core technical competencies for deploying robotics models and post training them for specific tasks start to diffuse, I expect this to become very common.

Perhaps by 2030, you will order a set of robots for your factory and hire a consultant to train them for a couple days if you’ve never done so yourself. Eventually, though, this seems like an economic efficiency that will largely be done away with. You don’t usually hire an external contractor to integrate an LLM into your workflows; robots should be no different.

I started this post with an anecdote about horses being replaced by the automobile, but so far I really haven’t discussed what, exactly, the horses are that are being replaced. These metaphorical horses aren’t people, not exactly, but they are jobs currently done by people. Software engineering, for example, is clearly harder to break into now than it used to be. An individual experienced programmer is so much more productive with AI as a “force multiplier,” which means that less people are necessary to build and deploy complex software products.

We should expect to see similar trends across basically every industry: more highly paid experts being far more productive, building more things, but each individual set of expertise being basically priceless, with robots handling more and more of the easily-replaceable labor. We should not fear this: in the developed world, working hours have largely been on a downward trend for decades, and it seems likely this will continue. Many human jobs will amount to handling the long-term planning and coordination that AI and robots seem persistently bad at, and intervening when they fail; but this should make for comparatively easy and low-stress work for a great percentage of the workforce.

Reasons for Doubt

It’s strange to look back at my own predictions for robotics from the end of 2024. Back then, one of my chief concerns was how to build scalable “world representations” for long horizon reasoning. This remains a concern to me; I’ve honestly seen basically no significant progress in this space. There are impressive 3d world models now, like that of World Labs, but these are all generally creating single scenes, not modeling their evolution in response to new sensor data. For true embodied general intelligence, we still need to address these fundamental questions about how robots will represent their knowledge of the world over time.

In some ways, it’s a good thing: the blocker for deploying long horizon reasoning has never been that long horizon reasoning is all that hard, it’s always been that execution is hard and that robots break (see, again, the Sankey diagram above from OK-Robot).

This also will still require lots of funding. Fortunately, it seems that many investors and billionaires with deep pockets — Jeff Bezos, Elon Musk, and others — are “all in” on artificial intelligence and robotics. Very likely, the money will not run out. But if it does, all this could come to a premature end.

And finally I worry about the closing-off of ecosystems. Open robotics innovation right now is championed, largely, by Physical Intelligence and NVIDIA, with some great recent entrants from Amazon FAR. While many of these problems are being solved, we still need new ideas and open dialogue to solve them — if all our doors and windows close, it’s possible the field stagnates and nothing gets accomplished.

Final Thoughts

With robot hardware getting so much better, with methods becoming mature and real-world results and long-running demos becoming somewhat common, I’ve never been more optimistic about what robotics will be capable of.

Part of the point of writing this is to point out how fast things have changed. And this is not unprecedented! I started this blog post with an anecdote about the horse being replaced by the automobile over just a couple decades. It’s not unreasonable to think that — in a lot of ways — we’re heading for such a moment. Not next year, not in two years, but over the next decade? It seems inevitable.

Basically all robotics problems get easier at scale: hardware related, deployment related, and data related issues all become much more tractable as soon as there are a lot of robots out there in the world.

I believe very strongly that we’re on a great trend, that these technologies will diffuse through society over the next 5-10 years, and that the future bright. But we’re headed for a very different world than the one we live in now, and that’s something we’ll also need to wrestle with over the coming years.

The future of industrial production is automated. From our ports to trucking and last-mile delivery, robots are now involved in nearly every part of bringing people the products that they want. And yet in a lot of cases the production of these products themselves is not yet automated; they are still built by specialized human workers. Even in wealthy, developed countries like the United States many things that seem as if they should be automated are not.

Take the example of Nike trying to near-shore production of shoes from Vietnam to Guadalajara, Mexico, as described in this Wall Street Journal article by Jon Emont. Shoe manufacturing relies on an army of skilled workers to perform fine manual labor, stitching and gluing a very wide variety of shoes together.

Ultimately, this effort was unsuccessful; Nike ended up closing the facility, and most shoes are still made by hand as of the writing of this article. And Nike isn’t the only example; small manufacturers in the United States rarely employ automation, far less than in near-peer countries like China or Germany.

There are a variety of reasons that automation seems to be taking off at different rates, ranging from technical to economic. Let’s go over some of the whys here: why automation is hard, where it fails, and why it seems to be moving faster in some areas than others.

If you like this blog, please consider subscribing, liking, or leaving a comment. Liking and subscribing helps others find these posts!

Subscribe now

When Automation Works Well

Industrial automation actually works very well within its narrow operational constraints. So-called dark factories have existed since 2001, when FANUC began lights-out operations at its flagship facility, with robots producing other robots wholly in the dark for up to 30 days at a time, completely unmanned.

These dark factories have become an iconic feature of Chinese industry, with giants like Xiaomi and BYD increasingly employing them to mass-produce products like smartphones and cars.

But all of these products have something in common: with smartphones, or cars, or industrial robots, you might be building millions of units of product with very few variations. This justifies the larger up-front cost of traditional automation, which involves carefully designing assembly lines, planning robot placement, and even planning individual movements the robots will be making in perfect synchrony months ahead of time.

This is, to say the least, an incredibly expensive undertaking. It’s the work of systems integrators, companies that focus on building a particular class of robotics automation solutions. They will plan camera placements, write code, hook up sensors, design custom tools or parts — whatever is necessary to produce the perfect fully-autonomous production line.

Reasons People Skip Automation

Above: video of a small Chinese machine shop by Marco Castelli on X

Benjamin Gibbs of Ready Robotics wrote a thread on X a while ago, listing reasons why you don’t see more robots deployed, especially looking at small and medium-sized enterprises (SMEs). In short:

• Skepticism: the idea that many people know someone or have themselves tried to deploy a new industrial robot, and have not gotten a good return on their investment.

• Opportunity cost: for a small, low-margin producer, spending $50,000 on a new robot makes a lot less sense than buying a new tool (also very expensive!) which can open up new revenue streams for the company

• Software complexity: note that this is not the complexity of programming the robot, but of integrating various vision and safety systems, connecting to industrial PLCs, and so on. Each integration can be a massive undertaking with many different programming languages and (usually poorly documented) software packages involved.

• Tooling design: industrial robots, as they are currently used in the United States, require very specialized end effectors to be useful. There is still no “universal” robot tooling; you cannot just order these parts off the shelf. Companies like Right Hand Robotics which once aimed to build more broadly-useful grippers and tools have always ended up narrowing their ambitions substantially.

• Parts presentation: basically, the “art” of building a reliable input system for parts to arrive at the robot; almost every one of these has historically been custom-made and is therefore wildly expensive.

• Electrical complexity: because there are so many custom parts and sensors, a shop looking to automate will need a custom electrical panel. They’ll need someone with wiring experience that they almost certainly lack in-house.

Let’s go back to the Nike example above. Their hope was that they could reproduce successes in building microprocessors, but applied to this new domain. Shoes are representative of a lot of the issues that robotics problems struggle with:

• The wide variety of shoes produced, all with subtle differences, increases the human effort necessary to handle the full range of products during the process of automation systems integration

• Similarly, the fact that shoes are made out of deformable materials means that the number of special parts and designs needed to fixture a shoe properly to grasp it 99.99% of the time and automate a full production line is extremely high

• Safety sensors and integration requirements exist for partially automating production, which would not exist for a setup that’s not automated

• Integrating all the specialized hardware used by a traditional systems integrator will be nearly impossible.

And the problem gets worse.

Automation in the United States

In 2021, China overtook the United States for number of robots deployed in manufacturing. By 2025, they’ve also overtaken anyone else, becoming the leading user of industrial robots in the world. This is not because of some technical edge — the bleeding edge of technology exemplified by companies like Physical Intelligence remains solidly American — but because of a complex ecosystem that enables production and deployment of robots economically and at scale.

As Andreesen-Horowitz noted in their report on robotics automation:

There are no “dark factories” in the United States. The closest that we have is Tesla’s Gigafactory Nevada, which is 90 percent automated. No other major manufacturer comes close.

Deployment costs and “cost disease” are endemic in many areas of American industry, and manufacturing appears to be no exception. As a result of these deployment hurdles, we often see that companies deploy robots only because they are forced to, usually by labor shortages.

It’s also noteworthy that, as mentioned above, current industrial automation requires a large number of custom parts, which means that the best returns will only appear at the very largest manufacturing scales. Without a robust ecosystem for custom systems integration work and a broad depth of expertise in the market, integration costs will remain extremely high.

So, in a way, it could be argued that we don’t use many robots, because costs are high, and costs are high because we don’t use many robots. This is a place where government action has worked in China, where there are often 10-20% up-front cash subsidies available for manufacturers. While the USA has similar incentives American subsidies take the form of tax breaks (such as Section 179), which necessitate extra legal fees, overhead, and for the company to take that extra cost burden ahead of time anyway only to be “paid back” later. It’s perhaps a less efficient way for the government to spend the same amount of money.

The Future

In the end, we have two types of problems: shortcomings of current technology and shortcomings of the broader economy that make automation less tractable in the united states. Aggressive competition in manufacturing has led to better returns on scale, both internally to companies and to the country as a whole which benefits from robust supply chains, expertise, and labor markets.

But technology, here, might be a way out. The factory of the future probably does not use the wide range of specialized sensors and fixtures that Ben Gibbs mentioned in his thread. Modern vision-based AI systems are pretty good at handling deformable materials like clothes. Companies like Dyna and Physical Intelligence have shown dual-armed mobile platforms that are both reasonably affordable and capable of performing an extremely wide variety of useful tasks to a high degree of reliability, if not particularly quickly (yet).

The difference here is massive. These newer, lower-cost robots are safer to be around, they’re cheaper to replace, and they don’t need the massive diversity of custom sensors or fixtures to work properly. Instead, because they work by seeing the world like humans do, they can be placed in a production line more or less the way that humans are, with only a fairly labor-intensive bringup process to teach them a new skill. But note that this bringup process is still less labor intensive than the current arduous process undertaken by the systems integrators we discussed above!

Perhaps in the farther future, we’ll see things like the MicroFactory take off: a bunch of cheap, modular robotic cells designed to be deployed in a controlled work cell, so that you could easily parallelize reinforcement learning training and deploy the robots at scale on a large production line. We’re already seeing companies like Standard Bots working to reinvent the formula.

Real technical challenges remain around speed, safety, and the general capability of these robots — as well as how to make their training more scalable and make the hardware more reliable. But the future of robotics manufacturing, it seems, will come from Silicon Valley and Austin, Texas, not the traditional manufacturing centers of the US, and it will be software-first.

If you liked this post, please like, share, and subscribe; or leave a comment with your thoughts below.

Subscribe now

Share

Leave a comment