Scholarlink

<p>In recent years, quadruped robots have been expected to play an active role in urgent missions such as search and rescue operations in disaster sites. However, sudden joint failures in such harsh environments remain a significant challenge. In particular, a <i>free-swinging joint failure</i> (<i>FSJF</i>), in which the actuator completely loses torque and the joint rotates freely under external forces and gravity, is a severe failure that makes posture stabilization extremely difficult. Despite its frequent occurrence in real-world environments, most previous studies have primarily focused on <i>locking failures</i>, leaving FSJFs insufficiently addressed. In this study, we focus on the scenario in which “any single joint suddenly fails” in a quadruped robot and develop fault-tolerant locomotion using deep reinforcement learning. Specifically, we incorporate the failure state as an input parameter and train a single model capable of handling FSJFs across all 12 joints. Simulation experiments demonstrated that the learned policy successfully avoided falls and maintained locomotion with a high success rate, even when FSJFs occurred at random timings during walking. Furthermore, we analyzed the post-failure locomotion behavior for each joint, providing insights that will contribute to the design of future fault-tolerant control strategies. These results highlight the importance of ensuring high reliability, particularly mission continuity for quadruped robots, and provide a useful foundation for real-world applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning a unified fault-tolerant gait policy for quadruped robots under single-joint sudden free-swinging failures

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Keisuke Matsuura,
  • Noritaka Sato

摘要

In recent years, quadruped robots have been expected to play an active role in urgent missions such as search and rescue operations in disaster sites. However, sudden joint failures in such harsh environments remain a significant challenge. In particular, a free-swinging joint failure (FSJF), in which the actuator completely loses torque and the joint rotates freely under external forces and gravity, is a severe failure that makes posture stabilization extremely difficult. Despite its frequent occurrence in real-world environments, most previous studies have primarily focused on locking failures, leaving FSJFs insufficiently addressed. In this study, we focus on the scenario in which “any single joint suddenly fails” in a quadruped robot and develop fault-tolerant locomotion using deep reinforcement learning. Specifically, we incorporate the failure state as an input parameter and train a single model capable of handling FSJFs across all 12 joints. Simulation experiments demonstrated that the learned policy successfully avoided falls and maintained locomotion with a high success rate, even when FSJFs occurred at random timings during walking. Furthermore, we analyzed the post-failure locomotion behavior for each joint, providing insights that will contribute to the design of future fault-tolerant control strategies. These results highlight the importance of ensuring high reliability, particularly mission continuity for quadruped robots, and provide a useful foundation for real-world applications.