I am a final year robotics and AI PhD student in the Nuclear and Applied Robotics Group at UT Austin, advised by Mitch Pryor. My current research focuses on robotic skill and behavior composition for modular robots. The first two years of my PhD focused on game theory, trajectory generation, and nonlinear optimization.

I co-founded geCKo Materials in 2021, a hardware manufacturing startup commercializing biomimetic industrial adhesives. We've recently raised ~$10M, built a team of 10+, and ship gMDA weekly. Our CEO and the team have been crushing it!

During my PhD, I spent summers at Cruise on the Behaviors team working on DriveGPT and on the Motion Planning & Control Analysis team working on performance metrics, and at GM working on visuomotor diffusion policies for grasping and manipulation. Before my PhD, I spent time at HP Labs in Palo Alto, CA as a Roboticist and Research Engineer, working across mobile robotics, AI/ML, 3D printing, and edge compute.

Education

PhD, Mechanical Engineering · UT Austin (Expected 2026)
Robotics & AI
Advisor: Mitch Pryor · Nuclear and Applied Robotics Group
MS, Mechanical Engineering · UCLA (2014)
Systems & Control
Advisor: Robert M'Closkey
BS, Mechanical Engineering · UCLA (2013)

Experience

geCKo Materials Campbell, CA | 2021 – Present
Co-Founder / Advisor / Board Member
  • Co-founded a biomimetic industrial adhesives company; contributed to $10M raise.
  • Scaled team to 10+ FTEs in a 10,000 sq. ft. manufacturing facility with weekly product shipments.
General Motors Mountain View, CA / Austin, TX | Summer 2026
Summer Researcher
  • Built a configurable scripted-policy API to generate synthetic visuomotor manipulation training data.
  • Trained diffusion policies for grasping and manipulating GM parts.
  • Built a custom evaluation harness to benchmark grasping and manipulation performance.
Cruise LLC San Francisco, CA | 2023 – 2024
Behaviors PhD Intern (Summer 2024) · Motion Planning & Control PhD Intern (Summer 2023)
  • MLE on DriveGPT; autoregressive decoder in a large driving model for NPC vehicle trajectory prediction.
  • Kinematic and dynamic metrics for trajectory generation analysis.
HP Labs (AI & Emerging Compute Lab) Palo Alto, CA | 2016 – 2020
Roboticist & Research Engineer
  • Designed PoCs for compute-constrained mobile robots with off-board ML inference.
  • Tried, to no avail, to convince HP executives to move into robotics.
  • Robotic cleaning solutions for 3D Print.
HP Inc. (Hardware R&D, 3D Print) San Diego, CA | 2014 – 2016
Control System and Servo Engineer
  • Built FDM 3D printer prototype from first principles; MIMO control laws for xyz, extruder, and temperature axes.

Publications

Contract-Grounded Behavior Tree Synthesis via Coding Agents
Contract-Grounded Behavior Tree Synthesis via Coding Agents
J Salfity, B Anderson, M Pryor
In submission pdf Project Webpage
A contract-grounded architecture for synthesizing deployable robot behavior trees from natural language, in which a coding agent queries a robot-side MCP server for an explicit contract of skills, operators, and composition templates before generating a validated, executable behavior tree. Non-expert operators can issue commands without knowledge of robot implementation details, while a runtime validation gate enforces correctness. Evaluated across 110 simulated PyRoboSim tasks and 14 tasks on a physical Husarion Panther robot with both closed (Sonnet 4.6) and smaller open-source (Gemma4:31b) LLMs, showing near-perfect BT validation, strong task success, and successful transfer to physical hardware running an opaque Nav2 stack.
Deployment Is Not Destiny
Deployment Is Not Destiny: Robot Recomposition in the Field with Unseen Software, Hardware, and Compute Payloads
S Swanbeck*, J Salfity*, J Gunawan, C Van Sice, B Anderson, M Pryor
In submission pdf Project Webpage
A plug-and-play framework for runtime recomposition of modular robots, enabling non-expert users to integrate unseen software, hardware, and compute payloads in the field within minutes. Demonstrated in disaster response scenarios at an operational nuclear reactor facility and in thermal-guided search.
Limited Linguistic Diversity
Limited Linguistic Diversity in Embodied AI Datasets
S Wanna*, A Luhtaru*, J Salfity, R Barron, J Moore, C Matuszek, M Pryor
Association for Computational Linguistics (ACL), July 2026 paper
A systematic audit of widely used VLA corpora showing that many datasets rely on highly repetitive, template-like commands with limited structural variation. We quantify instruction language across lexical variety, duplication, semantic similarity, and syntactic complexity to support more principled dataset selection and targeted augmentation strategies.
No Reboot Required
No Reboot Required: Real-Time Modular Robot Payload Integration
J Salfity, C Van Sice, B Chen, S Swanbeck, B Anderson, M Pryor
IEEE International Conference on Ubiquitous Robotics, July 2025 paper
A runtime integration system enabling modular robots to incorporate new hardware and software payloads during live operation, eliminating the need for system reboots and enabling continuous deployment.
Temporal and Semantic Metrics
Temporal and Semantic Evaluation Metrics for Foundation Models in Post-Hoc Analysis of Robotic Subtasks
J Salfity, S Wanna, M Choi, M Pryor
IEEE International Conference on Ubiquitous Robotics Workshop on Virtual Environment-Based Synthetic Data Generation for Robotic Training, July 2025 pdf
Temporal and semantic metrics for evaluating LLMs & VLMs ability to decompose robot trajectories into subtasks, benchmarking FM outputs against ground-truth subtask labels.
Inverse Matrix Games
Inverse Matrix Games with Unique Quantal Response Equilibrium
Yue Yu, J Salfity, D Fridovich-Keil, U Topcu
IEEE Control Systems Letters, 2022 paper
Inverse game problems in multiplayer matrix games under perception noise, inferring cost functions such that a desired joint strategy constitutes a unique quantal response equilibrium, with algorithms based on semidefinite programming and bilevel optimization.
Automated Decaking
Development of a Robotic System for Automated Decaking of 3D-Printed Parts
H Nguyen, N Adrian, JLX Yan, J Salfity, W Allen, QC Pham
IEEE International Conference on Robotics and Automation (ICRA), May 2020 paper
A robotic system using computer vision and manipulator control to automate the removal of residual powder from 3D-printed parts, enabling scalable post-processing for powder-bed fusion manufacturing.
ML Inference Offload
Cost-effective Machine Learning Inference Offload for Edge Computing
C Makaya, A Iyer, J Salfity, M Athreya, MA Lewis
HP Labs Technical Report, 2020 pdf
An offloading mechanism that leverages on-premises edge compute resources to handle ML inference for resource-constrained devices, reducing reliance on distant cloud servers. Demonstrated on person and object detection for mobile robots and video surveillance with improvements in accuracy and latency.

Patents

3D Printed Object Cleaning. J Salfity, W Allen, H Nguyen, N Adrian, JX Yan, QC Pham. US12311605B2 (Granted 2025)
Mobile Robots to Generate Reference Maps for Localization. J Salfity, D Murphy. US11703334B2 (Granted 2023)
Mobile Robots to Generate Occupancy Maps. J Salfity, D Murphy, W Allen. US11562524B2 (Granted 2023)
Transmission Rate Modification Based on Predicted Data. A Iyer, C Makaya, J Salfity. US2020/0353193A1 (Filed 2019)
Mobile Autonomous Fleet Control. W Allen, J Salfity. WO2020122953A1 (Filed 2018)
Modules of Three-Dimensional Printers. K Erickson, J Salfity, L Zhao. WO2020046267A1 (Filed 2018)
Control System for Mobile Robots. J Salfity, W Allen, H Horii. WO2019088990A1 (Filed 2017)
Correction of Filament Parameters. S Stodder, J Salfity, M Majette. WO2017086908A1 (Filed 2015)