Residual Learning on a Physics Pivoting Controller

Faster Search, and Where the Quasi-Static Model Breaks

Abstract

A parallel-jaw gripper can reorient a grasped rod by loosening its grip and letting gravity pivot it. In earlier work we built a model-based controller for this task that measures the rod angle from wrist motor current and steps the grip force down using a quasi-static friction model. That controller is accurate but slow: it searches for the slip threshold in small fixed steps, and each measurement costs a 10 s wrist sweep. Here we ask whether a learned residual on top of the controller makes it faster, at what cost in accuracy, and how it compares with a policy learned without the model. In MuJoCo simulation we trained linear policies with evolution strategies (1560 episodes per run, three seeds per method) that adjust each grip step and the release rate. On 40 held-out objects, residual policies cut the measurements per pivot from 18.0 to 9.3 and pivot time by 24%, with almost all of the saved steps in the search phase, but overshot the target more often (88% vs 95% of objects within 0.05 rad). Policies learned without the model were slower than the residual policies, and less accurate with noise-model sensing. Most overshoots happened on a single step: when the rod first slips it rotates in one jump (median 0.46 rad), which tended to be larger for larger force steps and faster releases, and which overshoots shallow targets. A guard that bounds the search step by the target angle and, once the rod moves, lets the residual be only more cautious than the model brought overshoots back to the model controller's level (93% within 0.05 rad) while keeping a 19% time saving; the guard was designed after we saw test results, so we confirmed it once on 40 fresh objects, where the guarded policies reached 98, 95 and 90% within 0.05 rad against 95% for the model controller and were 18% faster. With the full motor-current sensing pipeline the guarded residual was 32% faster than the model controller (3.73 vs 5.49 simulated minutes) at 97% vs 100% within 0.05 rad. One of three seeds stayed clearly worse with noise-model sensing, a release-rate cap added after training made every policy worse, and on heavier out-of-range objects the residual and model controllers degraded alike while pure RL degraded more. All results are from simulation.