• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (11): 191-200.doi: 10.3901/JME.260393

• 机器人及机构学 • 上一篇    

扫码分享

基于多评论家模仿学习的人形机器人运动技能生成方法

邵士博1, 林诚然2, 杜国锋1, 李俊年1, 曹政才2   

  1. 1. 北京化工大学信息科学与技术学院 北京 100029;
    2. 哈尔滨工业大学机器人技术与系统全国重点实验室 哈尔滨 150006
  • 收稿日期:2025-11-04 修回日期:2026-03-13 发布日期:2026-07-29
  • 作者简介:邵士博,男,1997年出生,博士研究生。主要研究方向为足式机器人运动控制。E-mail:shaosb97@gmail.com;曹政才(通信作者),男,1974年出生,博士,教授,博士研究生导师。主要研究方向为人工智能算法、机器人具身智能和人形机器人。E-mail:caozc@hit.edu.cn
  • 基金资助:
    黑龙江省自然科学基金(ZD2024E003)、国家自然科学基金(52535001)、航空科学基金(2025L015077001)和北京市自然科学基金(L243004)资助项目。

Whole-body Motion Skills Generation for Humanoid Robots Based on Multi-critic Actor Imitation Learning

SHAO Shibo1, LINChengran2, DU Guofeng1, LI Junnian1, CAO Zhengcai2   

  1. 1. College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029;
    2. State Key Laboratory of Robotics and Systems, Harbin Institute of Technology, Harbin 150006
  • Received:2025-11-04 Revised:2026-03-13 Published:2026-07-29

摘要: 人形机器人的运动能力是其实现复杂行为和执行多种任务的重要基础。现有运动控制方法在仿真环境中虽表现良好,但由于人形机器人具有高自由度、强耦合的特性,在面对现实世界的动力学差异、传感器噪声等挑战时,其运动技能的模仿精度与实物迁移稳定性均难以保证。为此,提出一种基于多评论家(Multi-critic actor)模仿学习的人形机器人运动技能学习方法。首先,设计基于动作复杂度指标的难度感知课程学习策略,使机器人能够从多样化动作数据中逐步学习复杂运动技能;其次,引入多评论家—演员网络结构,在提升对原始动捕数据模仿精度的同时,有效平衡动作模仿精确性与仿真到现实的迁移能力。实验结果表明,该方法在宇树G1机器人平台上实现了高精度动作模仿,并成功将仿真环境中学习到的运动技能零样本迁移至真实机器人,验证了所提方法在真实场景中的有效性与泛化能力。

关键词: 人形机器人, 多评论家模仿学习, 运动控制

Abstract: Whole-body motion skills are fundamental for humanoid robots to achieve complex behaviors and perform diverse tasks. Although existing motion control methods perform well in simulation, transferring them to the real world remains challenging. Humanoid robots typically have high degrees of freedom and strong dynamic coupling. As a result, accurate motion imitation and stable sim-to-real transfer are difficult to achieve under dynamics discrepancies and sensor noise. To address these issues, this research proposes a humanoid motion skill learning method based on multi-critic actor imitation learning. First, a difficulty-aware curriculum learning strategy based on a motion complexity metric is designed, enabling the robot to progressively learn complex motion skills from diverse motion datasets. Then, a multi-critic actor network architecture is introduced to improve the fidelity of imitation to motion capture data while balancing motion imitation accuracy and sim-to-real transfer capability. Experimental results show that the proposed method achieves high-precision motion imitation on the Unitree G1 humanoid robot. Motion skills learned in simulation are successfully transferred to the real world in a zero-shot setting. This demonstrates the effectiveness and generalization capability of the proposed method in real-world scenarios.

Key words: humanoid robots, multi-critic actor imitation learning, motion control

中图分类号: