Research
For a decision-making system to operate in the real world, its decisions must be interpretable to people and physically valid. We work along these axes.
Reliable Long-Horizon Behavior Synthesis
Keeping generated plans physically valid, beyond the horizon of the training data.
- Manifold-aware diffusion planning
- Trajectory stitching for long-horizon plans
- Detecting and refining infeasible plans
Physical AI
Decision making for embodied agents that act in the physical world.
- Manipulation and locomotion
- Unsupervised skill discovery via curriculum RL
Explainable AI
Interpreting the criteria behind a policy’s actions before execution.
- Feature attribution along faithful paths
- Interpretable deployment of learned policies
- Counterfactual activation editing
Where we are heading
We are extending the work in three directions.
Explainable Vision-Language-Action Models
Current VLA models produce infeasible or unsafe actions for ambiguous instructions such as “put the bowl in the microwave.” We analyze hidden states and attention to detect such hallucinations, with the goal of enabling a robot to ask for clarification when an instruction is ambiguous.
Inference-Time Scalability
Diffusion planners handle long-horizon planning well but do not improve with additional inference compute. Tree search exhibits inference-time scaling but lacks global coherence. We study the combination of the two toward System 2 reasoning.
Scaling to Manipulation and Multi-Agent Systems
Extending from navigation to manipulation tasks composed of sequential primitives, and further to environments with multiple interacting robots, where each agent must infer the intent of others and explain that inference to people.