ARC reasoning recipe lifts robot foundation models by up to 50 points with zero new data
imankitgoyal · x · 2026-10-09
Researchers introduced ARC, a reasoning recipe for robot foundation models that sets new SOTA on RoboLab-120 and MolmoSpaces without new robot data or foundation-scale training: up to +50.0 pp on RoboLab-Reasoning-50, +29.8 pp on RoboLab-120, and +27.2 pp on MolmoSpaces for π0.5 and Cosmos3-Nano-Policy.
The recipe has three ingredients: reasoning traces grounded in the robot's next action and its causal structure; a scalable automatic labeling pipeline (building ARC-Trace-DROID from DROID); and per-architecture fine-tuning and inference strategies so VLAs/WAMs can use traces for control. It offers an efficient complement to brute-force scaling.
Related event: Nvidia Unveils ARC: Robot Reasoning Success Boosted 5x Without New Data(3 posts)→
More from Embodied
- Hand-built robotic skin: AI asks to plug it into its head, reads touch in a minute — repligate · 2026-10-10
- Physical attention bias cuts cable-simulation prediction error by 15%+ — Avihai Giuili · 2026-10-10
- Wayve's AI Driver hits Turin roads in a Maserati Grecale just 4 weeks after integration — alexgkendall · 2026-10-10
- IROS 2026 Best Paper: humanoid robot learns tennis rallies from 5 hours of human motion data — qinzytech · 2026-10-10
- Tesla Teases Robotaxi "Floodgates Are Opening" as Expansion Signals Build — julianibarz · 2026-10-10
- Adaptive Sampling + Pure RL Yields Robot Policies Beyond Prior Locomotion Limits — jeffclune · 2026-10-10