LocoFormer: Generalist Locomotion via Long-context Adaptation

2025-09-28

AI Summary

Existing motion controllers rely heavily on manual parameter tuning and lack cross-morphology generalization. Method: We propose a universal locomotion control framework based on procedural robot architecture and strong domain randomization for large-scale reinforcement learning training, augmented by a long-context Transformer to enable cross-episode dynamics modeling and test-time adaptation—without requiring precise kinematic priors. Contribution/Results: The resulting policy achieves robust walking on unseen legged and wheeled robots. Experiments demonstrate seamless deployment across heterogeneous hardware platforms; the controller maintains stability under severe disturbances—including sudden payload changes, single-motor failure, and post-fall recovery—exhibiting strong generalization and emergent robustness. Notably, no morphology-specific fine-tuning or explicit dynamics modeling is required.

Technology Category

• Intelligent Robots: Behavior Learning & Control
• Machine Learning: Imitation Learning & Inverse Reinforcement Learning
• Planning, Routing, and Scheduling: Learning for Planning and Scheduling

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Studies of user behavior, including longitudinal effects of personalized systems

Abstract

Modern locomotion controllers are manually tuned for specific embodiments. We present LocoFormer, a generalist omni-bodied locomotion model that can control previously unseen legged and wheeled robots, even without precise knowledge of their kinematics. LocoFormer is able to adapt to changes in morphology and dynamics at test time. We find that two key choices enable adaptation. First, we train massive scale RL on procedurally generated robots with aggressive domain randomization. Second, in contrast to previous policies that are myopic with short context lengths, we extend context by orders of magnitude to span episode boundaries. We deploy the same LocoFormer to varied robots and show robust control even with large disturbances such as weight change and motor failures. In extreme scenarios, we see emergent adaptation across episodes, LocoFormer learns from falls in early episodes to improve control strategies in later ones. We believe that this simple, yet general recipe can be used to train foundation models for other robotic skills in the future. Videos at generalist-locomotion.github.io.

Problem

• Developing generalist locomotion controllers for diverse robot embodiments
• Enabling adaptation to unseen robot morphologies and dynamics
• Achieving robust control under disturbances and motor failures

Innovation

• Procedurally generated robots with aggressive domain randomization
• Extended context length spanning episode boundaries
• Adaptation across episodes learning from failures