Show-Harness presents a semantic interface for robot control that uses discrete action units, such as movement steps and gripper commands, which VLMs can reason over naturally. Embodiment-specific interpreters ground these units into physical motions, maintaining direct VLM responsibility for decisions.
The interface supports two modes: zero-shot deployment with frontier VLMs and low-cost fine-tuning of small VLMs (e.g., 2B-scale) using only a few GPU-hours. Additionally, GUMI extends the same action space for GUI-based demonstration collection without specialized hardware.
Extensive experiments demonstrate robust generalization across tasks, embodiments, and environments, outperforming representative agentic and VLA approaches. This suggests that a suitable action interface can unlock embodied capability from foundation VLMs without costly embodiment-specific pretraining.