SLAI Seminar 35th Session will be discussing the topic on "RoboClaw: Agent-Driven Self-Evolving Embodied Intelligence Systems ",  from 2:30pm-4pm, April 28 (Tuesday) at Room B401, online participation is welcome.

 (Tencent Meeting ID: 164-805-949)

报告主题：RoboClaw：智能体驱动的自进化具身智能系统

时间：2026年4月28日（周二）下午2:30-4:00

地点: 深圳河套学院B401教室

线上参与：腾讯会议号164-805-949

主讲嘉宾：穆尧教授

主持人：姬艳丽教授

讲者简介About the Speaker:

穆尧博士是上海交通大学长聘教轨助理教授，主要研究方向为多模态具身智能与机器人学习。穆教授博士毕业于香港大学计算机系，已在IJRR、RSS、NeurIPS、ICML、CVPR等顶级学术会议和期刊上发表论文50余篇。他曾担任ICLR等会议的领域主席，其研究成果在具身智能与机器人领域屡获殊荣并受到广泛认可。

Dr. Yao Mu is a tenure-track Assistant Professor at Shanghai Jiao Tong University. His research focuses on multimodal embodied intelligence and robot learning. He received his Ph.D. in Computer Science from The University of Hong Kong and has published more than 50 papers in leading venues including IJRR, RSS, NeurIPS, ICML, and CVPR. He has served as an Area Chair for conferences such as ICLR, and his work has received multiple awards and recognitions in embodied AI and robotics.

报告摘要：

本次报告将介绍 RoboClaw——一个面向可扩展的长时程机器人任务的智能体驱动框架，也是迈向自进化具身智能系统的一次新探索。RoboClaw不将数据收集、策略学习和任务执行视为孤立的阶段，而是将它们统一在由视觉语言模型（VLM）驱动的智能体之下，从而实现更一致的推理、策略编排和闭环改进。其核心理念是“纠缠动作对”（Entangled Action Pairs, EAP），它将任务行为与恢复动作相耦合，形成自复位循环，从而支持持续的在线策略数据收集与迭代优化。通过此设计，RoboClaw 减少了人工干预，提升了长时程操作中的鲁棒性，并为具身智能体通过交互实现持续学习、适应和演化指明了新范式。

Abstract:

This talk introduces RoboClaw, an agent-driven framework for scalable long-horizon robotic tasks and a new step toward self-evolving embodied intelligence systems. Rather than treating data collection, policy learning, and task execution as separate stages, RoboClaw unifies them under a single VLM-driven agent, enabling more consistent reasoning, policy orchestration, and closed-loop improvement. A key idea is Entangled Action Pairs (EAP), which couple task behaviors with recovery actions to form self-resetting loops for continuous on-policy data collection and iterative refinement. Through this design, RoboClaw reduces human intervention, improves robustness in long-horizon manipulation, and points to a new paradigm in which embodied agents can continuously learn, adapt, and evolve through interaction.