HexaShell:探索 VLM 智能体作为机器人编码智能体极限的薄封装
原标题:Paper page - HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
AI 摘要
研究团队提出 HexaShell,一个通过视觉和自写工具控制双臂机器人的编码智能体,在 RoboDojo 全部 42 项任务上取得 39.7% 成功率和 48.0 进度分,刷新该榜单最优成绩。该系统保持薄封装,仅将智能体命令转为机器人动作,智能体通过三路 RGB 相机与标定信息,在失败后读日志、定位原因并编写或修复小工具,以代码形式积累经验。其瓶颈在接触类任务而非推理:推动任务达 92.5%、记忆任务 65.0%,但紧密插入仅 8.3%。团队已开源目前演化出的工具。
正文节选
HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents Abstract HexaShell is a coding agent that controls a dual-arm robot via vision and self-written tools, achieving state-of-the-art results on RoboDojo but struggling with contact-heavy tasks. Agents built on vision language models (VLMs) now compete with learned policies on robot manipulation leaderboards. We ask how far a coding agent can go when it controls a robot instead of a computer, and where it stops.