Researchers have introduced Recursive Harness Distillation, a method aimed at helping vision-language-action (VLA) models handle manipulation tasks that require diagnosing failures and adapting behavior across changing tasks and environments.
According to the paper, VLA models offer broad manipulation capabilities but can struggle when execution demands failure diagnosis and behavioral adaptation. Strong agents can discover effective interventions by interacting with these policies, and the proposed method accumulates that experience as reusable guidance across agents.
In the approach, a strong agent distills its experience into a playbook for a light agent, then recursively refines the playbook using the light agent's execution feedback. The resulting playbook lets agents reuse accumulated intervention knowledge in new task instances without updating model parameters.
The paper reports that in real-world manipulation, the harness improves success from 37.3% to 64.0%. On SimplerEnv Bridge, the light agent with the playbook achieves 66.7% success, compared with 41.7% for the GR00T-only baseline, and outperforms the strong agent without a playbook. The same playbook also benefits the strong agent, which reaches 79.2% success.
The authors conclude that these results demonstrate the feasibility of harness distillation for robotics: intervention experience can be accumulated, refined through execution, and reused across agents to improve manipulation.