RoboDawn: VLM closed-loop robot control hits 73.6% one-shot, beating π0.5 zero-shot

Meng-Hao Guo · hf · 2026-09-22

RoboDawn exposes robot control to an agentic VLM via a compact set of discrete translation/rotation/gripper commands, enabling closed-loop observe-reason-act control, plus an in-context learning scheme grounded on a few demonstrations. On RoboTwin 2.0 C2R it reaches 53.2% zero-shot and 73.6% one-shot success, beating the strong trained baseline π0.5 (46.0%) and setting SOTA; RoboDojo improves from 35.67% to 47.17%. No task-specific robot training needed, and the same framework transfers to real Franka arms for block-in-basket and stacking tasks.

Original post →

More from Embodied

Embodied channel →