ARC claims more than 5× higher success on reasoning-heavy robot tasks

A post describes Nvidia presenting ARC, a method for adapting existing robot foundation models. The project says it labels existing demonstrations with explanations of why a robot’s next action is appropriate, then fine-tunes models to use them. It reports more than 5× higher success on reasoning-heavy tasks without changing model architecture. On RoboLab-120, it reports success rising from 36.8% to 48.8% for Cosmos3-Nano-Policy and from 28.0% to 45.3% for π0.5.