Learning to Face a Chair from Egocentric Depth
RL policy that orients a robot to face a chair using only an onboard depth camera, no external sensing.
Trained a reinforcement learning policy for a robot to turn and face the front of a chair using only an egocentric depth camera as input, with no external motion capture, localization, or other sensing of the chair. The behavior is fully learned end-to-end, with no hand-crafted heuristics.
Simulation result:
