Gemini Robotics On-Device 2: Evaluated on Dexmate

Abstract

Gemini Robotics On-Device is Google DeepMind's vision-language-action model built to run locally on the robot. It takes a text instruction, camera images, and the robot's joint state, and outputs robot actions. The model card for version 2 documents its training data, intended use, limitations, and safety evaluations. The model is available to a group of trusted testers.

The card's data-scaling evaluation uses robots the model saw only during post-training, and one of the two platforms is a Dexmate robot. As the training data per task increased, version 2 rose from 24.4% to 75.6% task success on it, while version 1 rose from 13.3% to 33.3%. DeepMind reports that version 2 learns new platforms faster and reaches a higher final success rate.

Citation

@misc{googledeepmind2026geminiroboticsondevice2,  title = {Gemini Robotics On-Device 2: Model Card},  author = {{Google DeepMind}},  year = {2026},  howpublished = {\url{https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/}}}

Platform

One of the two robots in the model card's data-scaling study is a Dexmate robot the model had not seen before post-training. On it, version 2 reached 75.6 percent task success as the training data per task grew, against 33.3 percent for version 1. Vega is available today as Vega 1 Pro and Vega 1 U. See all research published on Dexmate robots.

Related Research

• The Model Fills the Room. Hardware Builds the Room.
• ResFiT: Residual Off-Policy RL for Fine-Tuning Robot Policies