Researchers introduce ARC, a reasoning recipe that substantially improves zero-shot task performance of state-of-the-art robot foundation models (RFMs).

In this work, “zero-shot” refers to performing new tasks without task-specific fine-tuning; the reported gains were achieved without additional robot demonstrations or foundation-scale training.

ARC comprises three components: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. The authors report that effective reasoning traces are grounded in the robot’s next action and explain its causal structure — why the action is appropriate and what effect it should produce. They further report that such traces can be generated automatically from existing demonstrations and used to construct ARC-Trace-DROID from DROID without collecting new robot data.

The team adapted multiple model families. Examples include vision–language–action models (VLAs) such as π_{0.5} and whole‑arm manipulator policies (WAMs) such as Cosmos3‑Nano‑Policy. They applied fine-tuning and adjusted inference procedures tailored to each model’s architecture and capabilities. These interventions were intended to enable the models to use ARC traces for control.

Reported results are substantial on the studied benchmarks and models. The adapted models established a new state of the art on RoboLab-120 and MolmoSpaces. Gains reached up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC improved π_{0.5}’s task success by 82.2 percentage points. The authors describe these gains as unprecedented under the constraint of not adding robot demonstrations or doing foundation-scale training (phrased as “to our knowledge” in the report).