Robots operating in cluttered, unstructured environments often rely on external cameras for motion planning, which can be occluded or miscalibrated. A new system called ESCAPE instead uses egocentric sensing—depth sensors mounted on the robot—to generate collision-free trajectories.

ESCAPE aggregates sparse measurements from distributed depth sensors into a structured point cloud that highlights obstacles. A flow-matching generative policy then produces safe trajectories, trained on a dataset of over 100,000 safety-critical trajectories with joint optimization of goal-reaching and collision objectives.

The policy is trained entirely in simulation and transfers zero-shot to physical hardware without real-world fine-tuning. In both simulation and real-world evaluations, ESCAPE achieved stable motion planning using only egocentric perception, outperforming baselines that rely on external views or voxelized planning.