Recursive Harness Distillation across Agents for Robot Manipulation
The paper reports a robot-manipulation playbook that lifts real-world success from 37.3% to 64.0% without changing model weights.
Recursive Harness Distillation has a stronger agent turn its intervention experience into guidance for a lighter agent, then refine that guidance from the lighter agent’s execution feedback. The playbook is meant to carry failure-diagnosis and adaptation knowledge into new task instances. On SimplerEnv Bridge, the light agent reaches 66.7% success with the playbook, versus 41.7% for the GR00T-only baseline. The same playbook also helps the strong agent, which reaches 79.2% success. HF Daily Papers' note
Recursive Harness Distillation has a stronger agent turn its intervention experience into guidance for a lighter agent, then refine that guidance from the lighter agent’s execution feedback. The playbook is meant to carry failure-diagnosis and adaptation knowledge into new task instances. On SimplerEnv Bridge, the light agent reaches 66.7% success with the playbook, versus 41.7% for the GR00T-only baseline. The same playbook also helps the strong agent, which reaches 79.2% success. HF Daily Papers' note
score 4