RaC

A method that scales long-horizon task data by having a human take over right before failure, first recovering and then correcting.

RaC was released by Zheyuan Hu, Zackory Erickson, Aviral Kumar, and colleagues at Carnegie Mellon University in September 2025. The authors found that on contact-rich, long-horizon tasks involving deformable objects, simply piling on more expert demonstrations makes imitation-learning success rates plateau, because expert data almost never shows what to do after something goes wrong. RaC adds a human-in-the-loop fine-tuning stage after imitation-learning pretraining: while the policy is running, an operator takes over just before it fails, first steering the robot back to a familiar, in-distribution state (recovery), then demonstrating how to complete the current sub-task (correction), and ending the episode there. Across three real-robot bimanual tasks — hanging a shirt, sealing a food-storage container lid, and packing a takeout box — plus one simulated assembly task, RaC beats the previous best methods using roughly a tenth of the collection time and samples, and success rate scales roughly linearly with the number of recovery interventions performed during rollouts.

ExampleWhile hanging a shirt, the policy is about to hang it crookedly on the hanger; the operator takes over, first moves the arm back to a normal hanger-holding pose, then demonstrates hanging it correctly, and this intervention data is added to the fine-tuning set.

Also called

Recovery and Correction, RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction

Related

Recovery and Correction Data · Human-in-the-Loop · Human Intervention Data · Long-horizon Task · Compounding Error · HIL-SERL

Sources

RaC (arXiv 2509.07953)
RaC project page

As of

2025-09