ARC: A Reasoning Recipe for Robot Foundation Models

Forum topic · 小凯 · 2026-10-12

Summary

ARC is a reasoning recipe that substantially improves the zero-shot task performance of existing state-of-the-art robot foundation models (RFMs) without relying on larger models, more demonstrations, or costly large-scale training. The recipe has three ingredients: reasoning traces grounded in the robot's next action and its causal structure; a scalable automatic labeling pipeline that generates these traces from existing demonstrations (ARC-Trace-DROID, built from DROID without new robot data); and architecture-specific fine-tuning and inference strategies that let pretrained models such as the vision-language-action model pi-0.5 and the world action model Cosmos3-Nano-Policy use the traces for control. On RoboLab-120 and MolmoSpaces, the adapted models set new state-of-the-art results, with gains up to 50 percentage points on RoboLab-Reasoning-50. On real hardware, ARC improved pi-0.5's task success rate by 82.2 percentage points. The paper (arXiv:2610.12386) shows reasoning traces are a highly efficient complementary path to scaling robot policies.

Paper Overview

Field: Machine Learning Authors: Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss, Zhaoshuo Li, Fabio Ramos, Ankit Goyal, Jenai Xuning Yang Published: 2026-10-08 arXiv: 2610.12386

Summary

The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. This paper demonstrates 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. The authors call this recipe ARC.

ARC consists of three key ingredients:

1. Reasoning traces — Effective 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. 2. Scalable automatic labeling pipeline — These traces can be generated automatically from existing demonstrations. The authors built ARC-Trace-DROID from DROID without collecting any new robot data. 3. Adaptation strategy — State-of-the-art vision-language-action models (e.g., π₀.₅) and world action models (e.g., Cosmos3-Nano-Policy) learn to use these traces for control, with fine-tuning and inference tailored to each architecture.

Results

New state-of-the-art on RoboLab-120 and MolmoSpaces.
Gains of up to 50 percentage points on RoboLab-Reasoning-50.
On real hardware, ARC improved π₀.₅'s task success rate by 82.2 percentage points.

Original Abstract (excerpt)

> 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...

---

*Auto-collected on 2026-10-12*

ARC: A Reasoning Recipe for Robot Foundation Models

Summary

Paper Overview

Summary

Results

Original Abstract (excerpt)

Tags