Zero-shot VLM beats robot-trained π0.5 on RoboTwin benchmark

An off-the-shelf vision-language model, given only a compact set of discrete translation, rotation, and gripper commands, hit a 53.2% success rate on the RoboTwin 2.0 C2R manipulation benchmark, above the 46.0% scored by π0.5, a policy trained on the benchmark's robot data. The finding is reported in an arXiv preprint introducing RoboDawn, described by its authors as "a human-intuitive interface that exposes robotic control to an agentic VLM."

The setup is a closed loop: the model observes the visual state, reasons about the next action, executes it, and adapts based on what it sees. No robot-specific training. A single in-context demonstration lifts the same benchmark to 73.6%. On RoboDojo, numbers move from 35.67% zero-shot to 47.17% one-shot.

The authors frame the exercise around a broader question. "Can the intelligence of vision-language models (VLMs) similarly generalize from the digital world to the physical world for robotic control?" Their answer, in the abstract's own words: "RoboDawn achieves strong performance without task-specific robot training," and one demonstration "further yields substantial performance gains and establishes state-of-the-art (SOTA) results."

Beyond simulation, the paper reports that "the same framework also transfers to real-world robots, performing block-in-basket and block stacking on Franka." The abstract publishes no per-task real-world success rates, and it does not name the specific VLM sitting behind the closed-loop agent.