空间控制技术与应用2025，Vol.51Issue(2)：28-40,13.DOI:10.3969/j.issn.1674-1579.2025.02.003
多技能人形机器人全身运动策略智能生成方法
Whole-Body Motion Strategy Intelligent Generation Method for Multi-Skilled Humanoid Robots
张玲俊 1汤亮 2刘磊2
作者信息
- 1. 北京控制工程研究所,北京 100094
- 2. 北京控制工程研究所,北京 100094||空间智能控制技术全国重点实验室,北京 100094
- 折叠
摘要
Abstract
To address the challenge of enabling humanoid robots to acquire diverse whole-body motion skills under a single policy model while ensuring high-quality motion execution and smooth transitions between skills,a novel efficient multi-skill imitation learning method termed single model imitation learning for multi-skill efficiency(SMILE)is proposed.SMILE integrates goal-conditioned reinforcement learning(GCRL)and generative adversarial imitation learning(GAIL),introducing preference-based rewards tailored to the distinct motion characteristics of diverse skills,thereby mitigating the risk of convergence toward suboptimal policies.Furthermore,an adaptive failure-frequency-based priority sampling strategy is adopted,increasing the sampling probability of challenging samples to enhance learning efficiency and overall performance.Simulation results demonstrate that SMILE facilitates humanoid robots in performing diverse human-like whole-body motions,including standing,squatting,walking,obstacle jumping,stooping down for detailed inspection,and object picking.The trained policy achieves smooth skill transitions with an overall success rate of 93.33%,providing novel insights into multi-skill imitation learning for humanoid robots.关键词
人形机器人/强化学习/模仿学习/奖励塑造Key words
humanoid robots/reinforcement learning/imitation learning/reward shaping引用本文复制引用
张玲俊,汤亮,刘磊..多技能人形机器人全身运动策略智能生成方法[J].空间控制技术与应用,2025,51(2):28-40,13.
