What's this about?
Show-Harness gives VLMs compact semantic actions that embodiment-specific interpreters turn into local robot commands. It supports zero-shot closed models and small open models adapted with a few GPU-hours.
What exactly is a “semantic action unit”?
It is a meaningful command such as move, grasp or release, positioned between free-form language and motor-level control. The VLM selects intent while deterministic code grounds it for a particular robot.
Why is this worth noting?
The paper led today's HF list at 25 upvotes and reports generalization across tasks, robots and environments. It argues interface design can unlock capability without costly embodiment-specific pretraining.