ARC: A Reasoning Recipe for Robot Foundation Models
ARC boosts existing robot foundation models by adding action-grounded reasoning traces, without new robot demonstrations or foundation-scale training.
The paper says ARC has three parts: reasoning traces tied to the robot’s next action, an automatic labeling pipeline, and model-specific adaptation for control. The authors generate ARC-Trace-DROID from existing DROID demonstrations rather than collecting fresh robot data. They report new state-of-the-art results on RoboLab-120 and MolmoSpaces, including gains up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC raises π₀.₅ task success by 82.2 percentage points. ArXiv · AI/CL/LG's note
The paper says ARC has three parts: reasoning traces tied to the robot’s next action, an automatic labeling pipeline, and model-specific adaptation for control. The authors generate ARC-Trace-DROID from existing DROID demonstrations rather than collecting fresh robot data. They report new state-of-the-art results on RoboLab-120 and MolmoSpaces, including gains up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC raises π₀.₅ task success by 82.2 percentage points. ArXiv · AI/CL/LG's note
score 5