Show‑Harness Connects General VLMs to Robots Through a Small Semantic Action Interface

The research maps actions selected by a vision-language model into deterministic, embodiment-specific robot commands and reuses the interface across hardware.

Put a small action language between model and machine

Show‑Harness, submitted September 9, addresses the cost of adapting a general vision-language model to each robot. A VLM selects discrete semantic action units it can reason about, such as grasping, moving or aligning. An embodiment-specific interpreter deterministically grounds them into local coordinates and actuator-level behavior.

The authors say one interface enables both zero-shot control with closed frontier VLMs and low-cost adaptation of smaller open VLMs. They also propose GUMI, a GUI manipulation interface through which people or agents can collect demonstrations across robot embodiments without specialized teleoperation hardware.

“Playing” a robot is not safe autonomous work

The authors report stronger generalization than representative agentic and VLA approaches across tested tasks and embodiments. This is a September 9 preprint without confirmed peer review, and its experiments do not establish behavior on unseen hardware, contact dynamics, sensor failures or operation around people.

A semantic interface may reduce model-porting cost, but deployment still needs collision limits, force and speed envelopes, emergency stops, isolated workspaces and human supervision as independent safety layers.

Primary source