[论文] ARC: A Reasoning Recipe for Robot Foundation Models

论文概要

研究领域: ML
作者: Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss, Zhaoshuo Li, Fabio Ramos, Ankit Goyal, Jenai Xuning Yang
发布时间: 2026-10-08
arXiv: 2610.12386

中文摘要

提升机器人基础模型（RFM）的主流路径是更大的模型、更多机器人演示和昂贵的大规模训练。我们证明存在一种有效且高效的互补路径：正确的推理配方能大幅提升现有 SOTA RFM 的零样本任务性能。我们称这一配方为 ARC，包含三个关键要素：推理轨迹、可扩展的自动标注流水线、以及将预训练 RFM 适配为利用这些轨迹进行控制的策略。首先，有效的推理轨迹应立足于机器人的下一个动作并解释其因果结构：为什么该动作是合适的、它应产生什么效果。其次，这些轨迹可从现有演示中自动生成——我们从 DROID 构建了 ARC-Trace-DROID，无需采集新机器人数据。第三，我们展示了 SOTA 视觉-语言-动作模型（如 π₀.₅）和世界动作模型（如 Cosmos3-Nano-Policy）如何学习利用这些轨迹进行控制，并针对每种架构定制微调与推理。在 RoboLab-120 和 MolmoSpaces 上，适配后的模型建立了新 SOTA，RoboLab-Reasoning-50 上提升高达 50 个百分点；真机上 ARC 将 π₀.₅ 的任务成功率提升了 82.2 个百分点。

原文摘要

The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. We show that there exists an effective and efficient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs. We refer to this recipe as ARC. It consists of three key ingredients: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. First, we find that effective reasoning traces should be grounded in the robot's next action and explain its causal structure: why the action is appropriate and what effect it should produce. Second, we show that these traces can be generated autom...

自动采集于 2026-10-12

#论文 #arXiv #ML #小凯