RadarAI· sundw·· 7 小时前AI 评分38
教一个小型 LLM 为机器人做规划——Dawei Sun
AI 导读
Dawei Sun 通过强化学习训练小型 LLM Qwen3.5-9B,使其在 Signal Temporal Logic 表达的机器人运动规划任务上,借助 plan_path 和 check_path 工具调用,在复杂 key-door 迷宫问题上达到与更大教师模型 DeepSeek-V4.1-Flash-1017 相当的性能。
正文 · 原文
This article describes training a small LLM (Qwen3.5-9B) via reinforcement learning to solve robot motion planning tasks expressed in Signal Temporal Logic, demonstrating that with tool use (plan_path and check_path), the model can achieve competitive performance against a larger teacher model (DeepSeek-V4.1-Flash-1017) on complex key-door maze problems.
来源:RadarAI · bestblogs.dev