Gemini Robotics On-Device 2: Model Card

Google DeepMind's model card for its on-device vision-language-action model reports data-scaling results on a Dexmate robot: 75.6% task success, up from 33.3% for the previous version.