Heterogeneous multi-agent learning in Isaac Lab
Brought heterogeneous-agent RL algorithms into NVIDIA Isaac Lab so teams of different robots can learn to work together in GPU-parallel physics simulation.
Selected work
Master's thesis: one robot, many agents
Split a Spot quadruped into cooperating agents, one per leg or joint group, and showed that HAPPO's sequential updates learn a better-coordinated gait at almost no cost in training speed.
Strawberry Robotics
Led the AI side of a two-armed strawberry harvesting robot: Isaac Sim environments, a kinematics simulator, 3D berry and stem perception, and robot learning.
Coordinating search with foundation models and MARL
Implemented the Multi-Agent Transformer in Isaac Sim to coordinate a team of different robots searching unstructured outdoor environments.
Other projects
-
Rover mapping and person detection
ROS 2 rover that maps its surroundings and points toward nearby people, with a YOLO model fine-tuned on infrared images running on a Jetson Nano.
-
Scene graphs for RL in VizDoom
Tested whether graph-structured scene information helps recurrent APPO agents in VizDoom.
-
Reinforcement learning for Hearts
Agents that learn the card game Hearts, with a custom Gym environment.
Experience and education
-
Co-founder & CTO, Strawberry Robotics
-
M.S. Computer Science, Utah State University
-
Software Developer, Western Ag Credit
Internal applications in JavaScript, C#, and SQL.
-
B.S. Computer Science, Utah State University