Recursive Harness Distillation lifts robot manipulation success from 37.3% to 64.0%

SeoulNatlUniv · hf · 2026-09-29

Researchers propose Recursive Harness Distillation: a strong agent distills intervention experience into a playbook, recursively refined via a lightweight agent's execution feedback — no parameter updates needed. Real-world manipulation success rises from 37.3% to 64.0%; on SimplerEnv Bridge, the light agent with playbook hits 66.7% vs 41.7% for GR00T-only, and the strong agent reaches 79.2%.

Original post →

More from Embodied

Embodied channel →