Preprint / Version 2

Residual Learning on a Physics Pivoting Controller

Faster Search, and Where the Quasi-Static Model Breaks

##article.authors##

DOI:

https://doi.org/10.31224/8542

Keywords:

Reinforcement learning, Gravitational Dynamics, motor current sensing, in-hand manipulation, Evolutionary systems, safety-critical systems, MuJoCo

Abstract

One of three seeds stayed clearly worse with noise-model sensing; on validation objects, a release-rate cap added after training made every seed at least as accurate as the unguarded policy; and on heavier out-of-range objects the residual and model controllers degraded alike while pure RL degraded more.

Downloads

Download data is not yet available.

Author Biography

Clement Aidoo, Canlabs

Researcher at Canlabs research team

Downloads

Posted

2026-10-15 — Updated on 2026-10-21

Versions

Version justification

Version 2 corrects a coding error in the rate-cap validation experiment (Section on the model-based guard). The cap improves rather than worsens accuracy. All other results are unchanged. See the correction note in the paper.