From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation

One-line takeaway

This paper targets history-dependent, long-horizon robot manipulation where task progress and past outcomes must be retained and errors compound; it organizes the high-level VLM planner into a self-evolving harness with memory on top of low-level action execution.

这篇论文研究依赖历史的长程机器人操作：任务进度和过往交互结果需要记住，部分可观测和执行错误会累积；作者把高层视觉语言模型组织成带记忆、能自我演化的执行框架来规划，配合底层动作执行。