Humanoid policies trained in simulation often look good when transferred to hardware, but the force interactions they learned do not survive the journey. The result is degraded torque efficiency, poor load regulation, and weaker robustness to impacts. A new preprint on arXiv proposes Contact-UAN, an unsupervised actuator network that acts as a residual world model to address this gap.
Rather than retraining the whole policy, Contact-UAN models the difference between simulated and real actuator dynamics. This residual correction is intended to let the existing policy retain the force-related behaviors that simulation taught it, while adapting to the physical actuator's actual response.
The paper is a single source, and the claims are the authors' own at this stage. No independent evaluation is included in the abstract, so the reported benefits should be read as the proposed method's intended outcomes rather than established results.