A Drawn Line Steered Three Robot Policies

RoboPrompt translates intuitive human guidance into an action draft, then lets the base diffusion or flow-matching policy refine that draft through its normal generation dynamics. The reusable steering module worked across Diffusion Policy, pi-0.5 and FastWAM without policy-specific architectural changes or steerability fine-tuning. The researchers then used corrected rollouts for DAgger-style improvement. After two or three rounds, average success increased 15.5% for pi-0.5 across three tasks and 21.3% across three policies on an Insert Bread task, while intervention counts fell by 44.0% and 81.9% in the respective comparisons. The tested settings are limited, but the interface makes occasional human intent a lightweight input rather than a full teleoperation session.