Gemini Robotics On-Device 2
A local vision-language-action model for robotics teams, helping robots turn visual and spoken instructions into actions such as navigation, grasping, and placement.
Tool overview
Adoption judgment: Gemini Robotics On-Device 2 is worth evaluating, but the supplied evidence is not enough to call it a production-proven solution. There are only three social sources: one X recommendation and two Zhihu answers, with eight combined upvotes and no comments on Zhihu. This supports attention around the model, not a claim that it is effective or reliable.
The available descriptions present it as a vision-language-action model running on the robot itself. Its intended role is to convert visual context and natural-language goals into robot actions while reducing network latency. A Zhihu account describes an Apollo demonstration in which the robot follows an instruction to put a watering can into a green box, walks to the table, avoids obstacles, picks it up, bends down, and places it. This suggests coverage of whole-body mobility and manipulation, but it is a reported demo rather than an independent test or benchmark.
The deployment threshold and cost remain unclear.