Switch: Learning Agile Skills Switching for Humanoid Robots
Yuen-Fui Lau, Qihan Zhao, Yinhuai WANG, Runyi Yu, Hok Wai Tsui, Qifeng Chen, Ping Tan
AI summary
Problem
Existing whole-body control methods struggle with flexible, stable transitions between distinct dynamic skills, often leading to low success rates, unnatural movements, or safety risks due to open-loop tracking and poorly modeled transitions.
Approach
The authors propose a hierarchical system that builds a Skill Graph from motion data to connect similar states across different skills, trains a unified whole-body tracking policy via reinforcement learning on this graph, and uses an online scheduler to plan feasible paths for real-time skill switching or recovery.
Key results
- Constructs a data-driven Skill Graph to augment training data with feasible cross-skill transitions
- Trains a unified whole-body tracking policy using buffer-aware imitation learning and progressive curricula
- Implements an online skill scheduler that performs real-time graph search for seamless switching and safety recovery
- Achieves high success rates and robust real-world deployment on a Unitree G1 robot for dynamic skills like Kung Fu and dance
Why it matters
It provides a practical, scalable framework for deploying agile, multi-skill humanoid robots in real-world environments where safety and seamless transitions are critical.
Abstract
Recent advancements in whole-body control through deep reinforcement learning have enabled humanoid robots to achieve remarkable progress in real-world chal- lenging locomotion skills. However, existing approaches often struggle with flexible transitions between distinct skills, cre- ating safety concerns and practical limitations. To address this challenge, we introduce a hierarchical multi-skill system, Switch, enabling seamless skill transitions at any moment. Our approach comprises three key components: (1) a Skill Graph (SG) that establishes potential cross-skill transitions based on kinematic similarity within multi-skill motion data, (2) a whole-body tracking policy trained on this skill graph through deep reinforcement learning, and (3) an online skill scheduler to drive the tracking policy for robust skill execution and smooth transitions. For skill switching or significant tracking deviations, the scheduler performs online graph search to find the optimal feasible path, which ensures efficient, stable, and real-time execution of diverse locomotion skills. Comprehensive experiments demonstrate that Switch empowers humanoid to execute agile skill transitions with high success rates while maintaining strong motion imitation performance.