One-shot Humanoid Whole-body Motion Learning

2025-10-29

AI Summary

Training locomotion policies for humanoid robots typically requires large-scale, high-fidelity motion-capture (MoCap) data, which is costly and labor-intensive to acquire. Method: This paper proposes a one-shot-driven whole-body motion learning framework that synthesizes diverse locomotion behaviors from only a single non-locomotive target motion clip and a generic walking dataset. It measures motion sequence similarity via order-preserving optimal transport and generates intermediate poses through geodesic interpolation on the pose manifold. Collision-avoidance optimization and motion retargeting are further integrated to enable end-to-end reinforcement learning policy training in simulation. Contribution/Results: The approach drastically reduces reliance on large MoCap corpora. Evaluated on the CMU MoCap dataset, it outperforms baseline methods across multiple metrics, achieving high-fidelity, diverse, and data-efficient whole-body motion control.

Technology Category

• Search and Optimization: Learning to Search
• Computer Vision: Motion & Tracking
• Intelligent Robots: Motion and Path Planning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems

Abstract

Whole-body humanoid motion represents a cornerstone challenge in robotics, integrating balance, coordination, and adaptability to enable human-like behaviors. However, existing methods typically require multiple training samples per motion category, rendering the collection of high-quality human motion datasets both labor-intensive and costly. To address this, we propose a novel approach that trains effective humanoid motion policies using only a single non-walking target motion sample alongside readily available walking motions. The core idea lies in leveraging order-preserving optimal transport to compute distances between walking and non-walking sequences, followed by interpolation along geodesics to generate new intermediate pose skeletons, which are then optimized for collision-free configurations and retargeted to the humanoid before integration into a simulated environment for policy training via reinforcement learning. Experimental evaluations on the CMU MoCap dataset demonstrate that our method consistently outperforms baselines, achieving superior performance across metrics. Code will be released upon acceptance.

Problem

• Learning humanoid whole-body motion from single demonstrations
• Reducing data collection costs for human motion imitation
• Generating collision-free policies via motion interpolation and optimization

Innovation

• One-shot learning with single motion sample
• Order-preserving optimal transport for motion distance
• Geodesic interpolation generates intermediate pose skeletons