Preprint 2026

ARC: A Reasoning Recipe for Robot Foundation Models

Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss, Zhaoshuo Li, Fabio Ramos, Ankit Goyal, Jenai Xuning Yang

arXiv:2610.12386

Paper Project BibTeX

ARC labels existing robot demonstrations with action-grounded causal reasoning traces, then adapts vision-language-action and world-action policies using those traces. — ARC combines action-grounded reasoning, automatic demonstration labeling, and adaptation of robot foundation models.

Abstract

ARC improves robot foundation models by teaching them to reason about the causes and expected effects of their next actions. It combines action-grounded reasoning traces, an automatic pipeline for labeling existing demonstrations, and model-specific adaptation strategies. The pipeline produces ARC-Trace-DROID without collecting additional robot data. ARC adapts both vision-language-action models, including π₀.₅, and world-action models, including Cosmos3-Nano-Policy. The adapted models achieve state-of-the-art results on RoboLab-120 and MolmoSpaces, with gains of up to 50 percentage points on RoboLab-Reasoning-50. Real-robot experiments improve π₀.₅ task success by 82.2 percentage points, without additional robot demonstrations or foundation-scale training.

BibTeX

@article{puthumanaillam2026arc,
  title={ARC: A Reasoning Recipe for Robot Foundation Models},
  author={Puthumanaillam, Gokul and Sun, Tao and Aljalbout, Elie and Reuss, Moritz and Li, Zhaoshuo and Ramos, Fabio and Goyal, Ankit and Yang, Jenai Xuning},
  journal={arXiv preprint arXiv:2610.12386},
  year={2026},
  doi={10.48550/arXiv.2610.12386},
  url={https://arxiv.org/abs/2610.12386}
}