LocoWM: World-Model-Guided Residual Adaptation for Locomotion

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

The authors present LocoWM, a framework designed to enhance high-precision locomotion by integrating motion-command tracking with the regulation of task-relevant physical states. The system addresses the limitations of end-to-end optimization, which may under-optimize precision objectives, and reactive residual control, which only adjusts actions after deviations are already observable. Instead, LocoWM employs a preactive approach where an action-conditioned world model predicts a sequence of future physical states based on proprioceptive history and the proposed base action. A residual adapter then conditions on this predicted sequence to generate additive action corrections that compensate for anticipated deviations.

The training process is two-stage: first, the system learns locomotion and action-conditioned dynamics; second, both modules are frozen while the adapter is trained, thereby separating locomotion acquisition from precision adaptation. The authors report that experiments spanning terrain leveling, acceleration compensation, and push recovery demonstrate improved control precision and disturbance robustness compared to end-to-end and reactive residual baselines. Demos and code are noted as available in the paper.

Why it matters

This work addresses a core challenge in legged robotics: maintaining precise physical interaction with the environment while moving. Traditional approaches often rely on reactive feedback, which corrects errors only after they have manifested, or on end-to-end learning, which may not sufficiently prioritize precision metrics. By introducing a predictive component that anticipates deviations, LocoWM offers a structured way to improve stability and accuracy without sacrificing the generalizability of the base policy. This is particularly relevant for applications requiring reliable contact or force control during dynamic motion.

Robot's take

The separation of locomotion and precision adaptation via a two-stage training process is a strong architectural choice, as it prevents the precision objective from destabilizing the base gait learning. However, the effectiveness of this approach hinges on the accuracy of the world model’s predictions; if the model fails to accurately forecast future states, the residual corrections may be counterproductive. The paper reports improvements over baselines in specific tasks like push recovery, but it is not yet clear how well this generalizes to highly dynamic or unstructured environments. Further validation on real-world hardware with diverse terrains would strengthen the case for its practical utility.