Multi-Embodiment Locomotion at Scale with extreme Embodiment Randomization

2025-09-02

AI Summary

This work addresses the zero-shot transfer challenge of general locomotion policies across morphologically diverse legged robots—including previously unseen real-world humanoid and quadruped platforms. We propose URMAv2, an embodied perception architecture that integrates extreme morphological randomization, performance-guided curriculum learning, and deep reinforcement learning to train a single unified policy across 50 heterogeneous legged morphologies. The framework enables efficient exploration of a million-scale morphological variation space and achieves, for the first time, end-to-end policy training across multiple robot morphologies with direct plug-and-play deployment on physical systems. Experiments demonstrate that the learned policy delivers high-performance walking control on unseen real-world robots, significantly improving generalization, robustness, and deployment efficiency. URMAv2 establishes a scalable new paradigm for universal locomotion control in embodied intelligence.

Technology Category

Intelligent Robots: Embodied AIHumans and AI: Human-Aware Planning and Behavior PredictionMachine Learning: Imitation Learning & Inverse Reinforcement Learning

Application Category

Responsible Web: Machine-in-the-loop, human agency and autonomySearch and Retrieval-Augmented AI: Multilingual and cross-lingual Web searchUser Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

We present a single, general locomotion policy trained on a diverse collection of 50 legged robots. By combining an improved embodiment-aware architecture (URMAv2) with a performance-based curriculum for extreme Embodiment Randomization, our policy learns to control millions of morphological variations. Our policy achieves zero-shot transfer to unseen real-world humanoid and quadruped robots.

Problem

• Control diverse legged robots with one policy
• Handle millions of morphological variations effectively
• Achieve zero-shot transfer to unseen real robots

Innovation

• Single policy for diverse legged robots
• Embodiment-aware architecture with extreme randomization
• Zero-shot transfer to unseen real-world robots