Compiling Short Sensor Histories into Frozen Robot Policies: A Closed-Loop Audit of Joint Encoder Dropout

Abstract

When a robot loses a proprioceptive channel, recent readings from its remaining sensors may keep a trained policy running. We fit a causal affine reconstructor from offline trajectories and fold it into the first layer of frozen PPO actors. Across Reacher, Hopper and Walker2d, eight-frame ridge reconstruction improves held-out mean return over one-frame ridge for all nine task and fault combinations. With clean angles available, finite differences approach the healthy reference, while removing angles as well as velocities leaves large losses on both locomotion tasks. A follow-up drops all joint velocities during motion and tests unseen angle noise, including temporally correlated errors. At 0.05 radian white noise, training ridge with noisy angles raises Walker2d's post-fault return from 310.9 to 1,164.5, with a cost under clean readings. Learning a correction to smoothed differencing helps Hopper in that condition. The results quantify how noise, retained information, and calibration data affect closed-loop repair.

Keywords