• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 168-178.doi: 10.3901/JME.260766

• 特邀专栏:人形机器人与具身智能 • 上一篇    

扫码分享

模块化双足机器人的行走过程关节失效容错机制研究

齐健, 赵思恺, 赵宁, 杨致远, 周楠林, 梁大伟, 罗航, 郑天骄, 赵杰, 朱延河   

  1. 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 150001
  • 收稿日期:2025-12-13 修回日期:2026-05-07 发布日期:2026-11-20
  • 作者简介:齐健,男,1996年出生,博士研究生。主要研究方向为模块化自重构机器人、强化学习运动控制。E-mail:20B908059@stu.hit.edu.cn;朱延河(通信作者),男,1975年出生,博士,教授,博士研究生导师。主要研究方向为模块化自重构机器人、可穿戴机器人、极端环境机器人、工业机器人等。E-mail:yhzhu@hit.edu.cn
  • 基金资助:
    国家自然科学基金资助项目(52025054,52435001,U25A20288)。

Research on Fault-tolerant Mechanisms for Joint Failures in the Walking Process of Modular Biped Robots

QI Jian, ZHAO Sikai, ZHAO Ning, YANG Zhiyuan, ZHOU Nanlin, LIANG Dawei, LUO Hang, ZHENG Tianjiao, ZHAO Jie, ZHU Yanhe   

  1. State Key Laboratory of Robotics and System, Harbin Institute of Technology, Harbin 150001
  • Received:2025-12-13 Revised:2026-05-07 Published:2026-11-20

摘要: 足式机器人在复杂恶劣环境中执行任务时,关节故障极易导致基于无故障模型设计的运动策略失效,机器人的运动性能显著下降甚至完全丧失。针对该问题,以模块化双足机器人为研究对象,提出一种面向关节故障的强化学习运动控制框架,使机器人仅通过单一控制策略即可同时适应健康状态与多种关节失效模式。框架采用非对称Actor-Critic结构,在Critic网络中引入包含故障状态的特权信息,并设计关节故障掩码以覆盖多样化的失效场景。训练过程中,提出渐进式故障课程学习方法,引导智能体先学习无故障基础运动,再逐步过渡到适应不同关节故障;同时引入故障率机制,协调健康与故障状态下的经验积累,避免遗忘已习得的运动技能。在IsaacGym仿真平台上进行实验,从速度追踪、抗干扰鲁棒性、轨迹跟踪等方面验证所提方法的有效性。结果表明:该方法在全部12个关节发生零力矩失效时均能保持稳定行走;与仅添加故障掩码的方法相比,速度追踪误差降低8.3%。进一步将故障模型扩展为力矩衰减和关节锁死故障,与基于师生策略的强化学习容错方法相比,所提方法在两种故障下的平均速度追踪误差分别降低23.5%和30.4%。研究揭示了模块化双足机器人不同关节对故障的敏感度差异,为提升机器人在关节失效条件下的自主生存能力提供了新的技术途径。

关键词: 强化学习, 课程学习, 模块化机器人, 容错, 关节故障

Abstract: When legged robots perform tasks in complex and harsh environments, joint failures can easily cause motion strategies designed under fault-free assumptions to become ineffective, leading to significant degradation or even complete loss of mobility. To address this issue, we focus on a modular biped robot and propose a reinforcement learning-based locomotion control framework for joint failures, enabling the robot to adapt to both healthy states and multiple joint failure modes using a single control policy. The framework employs an asymmetric Actor-Critic structure, incorporating privileged information of joint failure states in the Critic network and designing joint failure masks to cover diverse failure scenarios. During training, we introduce a progressive fault curriculum, guiding the agent to first learn fault-free locomotion and then gradually adapt to different joint failures. Additionally, a failure rate mechanism is introduced to coordinate experience accumulation between healthy and faulty states, preventing forgetting of previously learned skills. Experiments are conducted on the IsaacGym simulation platform to validate the proposed method in terms of velocity tracking, disturbance rejection robustness, and trajectory tracking. Results show that the method maintains stable walking even when all 12 joints experience zero-torque failure; compared with the approach using only fault masks, the velocity tracking error is reduced by 8.3%. Further, when extending the fault model to torque degradation and joint lock failures, the proposed method achieves 23.5% and 30.4% lower average velocity tracking errors, respectively, compared with a teacher-student reinforcement learning fault-tolerant method. The study reveals the varying sensitivity of different joints in the modular biped robot to failures, providing a new technical approach to enhance the robot’s autonomous survivability under joint failure conditions.

Key words: reinforcement learning, curriculum learning, modular robots, fault tolerance, joint faults

中图分类号: