I build infrastructure for robots that learn from the physical world.
My current work is RoboRun: an instrumented loop for robot data, RL environments, policy deployment, and replayable telemetry. It turns video, state, actions, and outcomes into training episodes and evidence without repeatedly rebuilding the underlying media.
- RoboRun — robot data infrastructure and runtime
- referee-lab — public capture-QA evaluators for robot-training video
- egosieve — manipulation-window discovery for egocentric video
- ros_tap — zero-configuration telemetry for ROS 1 and ROS 2
- dimos — an agentic operating system for physical space
capture → structure → execute → replay