Skip to content

论文HF Daily Papers · arXiv 2611.1600911 天前

JustRUA精简Harness使GPT-6在RoboDojo仿真得分提升70% ​

本文由 AI 根据公开内容整理,可能有误,请以原文为准:JustRUA: A Controlled Study of GPT-6 Robot-Use Agents(HF Daily Papers · arXiv 2611.16009)

核心内容 ​

针对GPT-6 Astra等模型驱动机械臂时因Harness设计与测试集差异导致性能归因困难的问题,研究团队在统一框架下解耦测试了Harness结构、动作接口、经验复用及场景几何四大变量。提出的JustRUA基于官方L3 Harness进行极简重构,剔除冗余模块,仅保留对性能贡献最大的核心组件,旨在验证基础工程优化对大模型机器人操控能力的上限提升作用。

实验表明,保留完整Episode历史并引入运动早停机制的轻量Harness效果显著;经验复用虽改变模型行为但非绝对正向收益;而场景几何信息(即使仅由RGB预测深度)对性能提升贡献最大。最终JustRUA在无任何学习策略、自改进循环或任务特定微调的情况下,将RoboDojo仿真基准得分从29.7大幅提升至50.3,证明了系统化Prompt与上下文工程的有效性。

关键看点 ​

  • 轻量Harness设计:保留完整交互历史并实施运动早停机制,显著提升GPT-6在RoboDojo上的操控成功率
  • 几何信息关键性:即便仅通过RGB预测深度,提供场景几何信息仍是提升Agent性能的最核心要素
  • 经验复用双刃剑:重用历史经验虽能改变模型决策模式,但在当前架构下并非无条件带来性能增益
  • 零训练性能跃升:无策略学习与微调,仅凭精简Harness即将RoboDojo得分从29.7提至50.3

工程与产业观察 ​

该研究揭示了具身智能落地中“上下文工程”的被低估价值。在端到端大模型尚存幻觉与推理瓶颈时,通过精细化设计Harness(如全量历史保持、几何先验注入)可作为低成本增效手段。这提示算法工程师在追求模型参数规模之外,应优先优化系统级Prompt结构与感知反馈链路,以最小算力成本挖掘基座模型的物理交互潜力。

英文摘要 (Abstract) ​

GPT-6 Astra can drive a robot arm directly through a fixed harness, and scores of such robot-use agents on RoboDojo have risen quickly since. Reported systems differ in harness and test subset and often change several designs at once, so it is hard to tell which choices produce the gains. We take these designs apart and test them in one framework, along the harness around the model, the action interface, the reuse of experience and the scene geometry given to the model. A lean harness that keeps the full episode history and stops each motion early goes a long way. Reusing experience changes how the model acts but is not a free lunch, and scene geometry helps most, even when depth is predicted from RGB alone. JustRUA starts from the official L3 harness and keeps only what matters most. Even with no learned policy, self-improvement loop or task-specific tuning, this simple harness raises the score on the RoboDojo simulation benchmark substantially, from 29.7 to 50.3.

Hugging Face Daily Papers 热度:1 upvotes 作者:Zefan Wang, Yufan Dang, Xuanwei Shi, Mingqi Gao, Cheng Yin, SikYuen Tam, Hanyu Liu, Yasheng Wang, Yaxi Lu, Songyang Zhang, Xin Cong, Yuan Yao

基于本页技术方向关联的整机参数、核心零部件、实测报告与同赛道选型横评。

外部来源

本页事实对应的公开页面,以原文为准。(点击将离开本站)

  1. HF Daily Papers · arXiv 2611.16009:JustRUA: A Controlled Study of GPT-6 Robot-Use Agents ↗访问 2026-11-13