Show-Harness Says Robots Don't Need a New Model, They Need a Better Interface
Show-Harness went up to 125 upvotes on HuggingFace papers, the top paper of the day, and the title is a shot at the entire vision-language-action field: just a VLM agent can play robots. No embodiment-specific pretraining, no VLA model, no teleop rig. The claim is that the reason foundation models are bad at robots is that we keep handing them the wrong interface.
The design is discrete semantic action units, meaning actions a VLM can actually reason about in language, paired with per-embodiment interpreters that translate those units into whatever joints and grippers the hardware has. The VLM stays inside the physical decision loop instead of being demoted to a high-level planner that hands off to a policy. Three things fall out of this: frontier closed models can do zero-shot robot control, small open models can be fine-tuned into it cheaply, and cross-embodiment transfer works because the semantic layer is shared. They also built GUMI, a GUI manipulation interface for collecting demonstrations without specialized teleoperation hardware, which quietly solves the most expensive part of robot data collection.
The paper reports beating comparable agentic and VLA paradigms across tasks, embodiments, and environments. The conclusion they draw is the one worth stealing: the right interface can unlock substantial embodied capability from foundation VLMs.
arXiv 2609.10522, submitted September 9, project page at https://showlab.github.io/Show-Harness
This is the harness thesis arriving in robotics. For a year the same argument has been playing out in coding and research agents, that the gains come from loop and interface design rather than weights, and every month someone shows a smaller model with a better harness beating a bigger one. Embodiment was supposed to be the exception, the place where you truly need new pretraining. If Show-Harness holds up under replication, it is not the exception either.
← Back to all articles
The design is discrete semantic action units, meaning actions a VLM can actually reason about in language, paired with per-embodiment interpreters that translate those units into whatever joints and grippers the hardware has. The VLM stays inside the physical decision loop instead of being demoted to a high-level planner that hands off to a policy. Three things fall out of this: frontier closed models can do zero-shot robot control, small open models can be fine-tuned into it cheaply, and cross-embodiment transfer works because the semantic layer is shared. They also built GUMI, a GUI manipulation interface for collecting demonstrations without specialized teleoperation hardware, which quietly solves the most expensive part of robot data collection.
The paper reports beating comparable agentic and VLA paradigms across tasks, embodiments, and environments. The conclusion they draw is the one worth stealing: the right interface can unlock substantial embodied capability from foundation VLMs.
arXiv 2609.10522, submitted September 9, project page at https://showlab.github.io/Show-Harness
This is the harness thesis arriving in robotics. For a year the same argument has been playing out in coding and research agents, that the gains come from loop and interface design rather than weights, and every month someone shows a smaller model with a better harness beating a bigger one. Embodiment was supposed to be the exception, the place where you truly need new pretraining. If Show-Harness holds up under replication, it is not the exception either.
Comments