Found first: a primary source the press has not covered yet.
A paper posted to arXiv on September 19, 2026 describes RoboDawn, a framework that exposes robotic control to an off-the-shelf vision-language model through a compact interface of discrete translation, rotation, and gripper commands. Running zero-shot, with no robot-specific training data, the VLM scores 53.2% task success on the RoboTwin 2.0 C2R benchmark against 46.0% for the trained π0.5 specialist policy. Adding a single in-context demonstration lifts the score to 73.6%.
What the source says
The work comes from nine authors led by Meng-Hao Guo and Yongming Rao; author affiliations were not listed on the abstract page. RoboDawn runs in a closed loop: the VLM observes visual state, reasons over it, issues commands, then adapts based on the resulting state. On RoboDojo, a second simulation benchmark, zero-shot performance is 35.67%, rising to 47.17% with one demonstration. The authors also run block-manipulation tasks on a physical Franka robot.
Why it matters
Robot manipulation research has operated on the assumption that task-specific training data is required for competitive performance. This paper challenges that directly: a general VLM with no robot training matches and exceeds a fine-tuned policy on the same benchmark. The one-shot result, 73.6%, shows that a single demonstration provides substantially more lift than an entire training run does relative to zero-shot. If the result generalizes, it changes the cost calculus for robot deployment, where data collection and curation are the primary bottleneck.