World Models

Force-Aware World Models for Contact-Rich Manipulation

A world model lets a robot imagine the outcome of an action before executing it. We are developing force-aware world models that predict future contact forces along with future camera frames, and output a target motion together with a stiffness for compliant execution. Contact force is estimated from motor current, without a dedicated force/torque sensor.

World Models for Neurovascular Intervention

We also build world models for robotic catheter navigation in stroke thrombectomy. Reinforcement learning agents trained in a digital twin of the vasculature, using pseudo X-ray images and virtual haptic feedback, learn to insert and rotate the guidewire and catheters. We distill these task-specific agents into a single world-model agent that performs multiple navigation tasks across different patient vasculatures and generalizes to unseen anatomies. During deployment, the robot can test a candidate motion in the simulator before executing it, checking for vessel-wall contact or instrument buckling, and then act on the next real X-ray frame.