Show-Harness Lets Frontier VLMs Drive Robots Zero-Shot

Vision-language models can drive robots zero-shot, a new arXiv preprint argues, provided their reasoning is routed through the right thin interface.

That interface is Show-Harness. Posted September 9, the paper describes it as "a compact semantic interface linking intent to action," exposing "discrete semantic action units that VLMs can naturally reason over," with embodiment-specific interpreters that "deterministically ground them into local robot actions." The VLM, the authors write, stays "directly responsible for fine-grained physical decisions."

Two demonstrations sit on top. One is zero-shot: closed-source frontier VLMs, driven straight through the harness with no training. The other is a cheap adaptation path: small open-source VLMs fine-tuned with "just a few GPU-hours."

The authors also introduce GUMI, a "GUI Manipulation Interface," which extends the same semantic action space to a graphical demo-collection tool so humans and agents can "play" robots "across embodiments without specialized teleoperation hardware."

The headline claim is that Show-Harness-equipped agents "generalize robustly across tasks, embodiments, and environments, outperforming representative agentic and VLA paradigms." The abstract names no baselines, lists no embodiments, and reports no per-task numbers. It lands in a busy stretch of humanoid-VLA releases our tracker has logged over the past week, including Berkeley's TANGO, a 29-DoF humanoid VLA on Unitree G1 the same day.