Current focus
Agentic robotics
Robots as agents, not just policies. A reasoning model plans, calls skills and tools (perception, navigation, manipulation primitives, ROS services), checks its own progress and recovers when things go wrong. Where do agent loops make robots more capable, and where do they just add latency?
- goal“put the red cup in the sink”
- planfind cup → grasp → place in sink
- calldetect(“red cup”)✓
- callgrasp(cup_1)✗ slipped
- verifyre-plan grasp from new view✓
- callplace(cup_1, sink)✓
Current focus
System 0 / 1 / 2
A layered control stack built on fast and slow thinking. The interesting questions are at the seams: what each layer learns, how often it runs and what data it needs.
- S2Reason & planVLM / LLM~1 Hz
- S1See & actvisuomotor policy (VLA)~10–100 Hz
- S0React & stabilizelow-level / whole-body control~1 kHz
Data scaling for robot learning
Which data actually moves the metrics, how performance scales with it, and where the next labeled hour should go.
Embodied foundation models
VLAs, world models, action chunking and flow-matching policies, and the data and infrastructure they really need.
Learning from humans
Egocentric and third-person human video as robot data, and how to bridge the embodiment gap.