Gemma 4 VLA Demo on Jetson Orin Nano Super

The story isn't just about a model running on a small board; it's about a vision-language-action model, which adds a third modality beyond seeing and talking — it can produce control outputs that drive physical actuators, making this relevant to robotics in a way that a standard VLM demo would not be. The Jetson Orin Nano Super sits at roughly the $250 price point, meaning the barrier to putting this capability inside a physical robot is now closer to hobbyist territory than research lab territory.

Google has been pushing multimodal capability across its model families at a steady pace. The Gemini 3.1 Flash TTS release from April 15 showed DeepMind adding expressive audio control, and the two Google Photos integrations covered around April 16 showed Gemini reaching into personal data for image generation. Those moves were all cloud-side or on-device in a phone context. This Gemma 4 VLA demo is a different branch of that effort: inference at the edge for embodied applications, not consumer software. The related coverage doesn't map cleanly onto robotics or embedded hardware, so this story largely belongs to a separate conversation about physical AI that Modelwire hasn't covered heavily yet.

Watch whether Google or a third-party robotics team publishes latency and reliability benchmarks for closed-loop control tasks (not just inference throughput) on this hardware within the next 60 days. Inference speed on a demo board means little until someone shows the model completing a manipulation task end-to-end without falling back to cloud.