Dexterous Manipulation and Humanoid Robots

Agentic Robots for Long-Horizon Tasks

Everyday tasks such as making coffee or packing an order are long sequences of skills. We are building agentic robots in which a task-level agent sequences skills and retries or re-plans when a skill fails. Our platforms include a bimanual OpenArm with wrist, head, and depth cameras, an OpenArm KER teleoperation device for collecting demonstrations, and Booster K1 humanoid robots. See Robots for details.

Force-Aware World Model Manipulation for Logistics

Many packing steps in e-commerce fulfillment still depend on human hands because items are deformable, cluttered, and highly varied. We are applying force-aware world models to packing, starting with bag-in-bag packaging: one arm holds a mailer bag open while the other inserts small polybags without tearing or crushing them.

Robots for Food Service: Coffee Making

In our coffee-making project, a bimanual OpenArm learns to prepare drip coffee, a sequence of steps that require precise, force-aware contact. We build a simulated coffee-making scene to collect demonstrations labeled with contact force.

Humanoid Locomotion Control

In our Science Robotics review with collaborators, we trace how humanoid locomotion control has evolved from classical model-based control and optimization, to reinforcement learning powered by large-scale simulation, and now to generative approaches such as foundation models and world models. We argue that these paradigms are converging toward physics-guided generative intelligence that combines physical modeling, constrained decision-making, and adaptation to uncertainty.