Google DeepMind has released Gemini Robotics 2, a family of three models that gives robots whole-body control, multi-minute task planning and the ability to coordinate with other robots. It’s the most significant update to Google’s robot “brain” since the original Gemini Robotics launched, and it lands as every humanoid company on earth is racing to build or buy exactly this kind of model.
The three models
- Gemini Robotics 2 (VLA) — the vision-language-action model that turns instructions and camera input into motor commands. Early access.
- Gemini Robotics-ER 2 — the embodied-reasoning model for planning and spatial understanding. Available now in Google AI Studio.
- On-device variant — runs locally on the robot for low-latency control. Early access.
Demos ran on Apptronik’s Apollo 2 humanoid, Franka’s Duo gripper and Sharpa Wave dexterous hands, and DeepMind says the models can adapt to a new robot body in hours rather than weeks. Reported success rates included 76.3% on a shelf-pickup task, 89.6% on insertion and 92% on unscrewing light bulbs.
Safety as a benchmark
Alongside the models, DeepMind introduced ASIMOV-Agentic, a safety benchmark for agentic robots. In demonstrations, robots refused commands judged unsafe and asked a human for help when uncertain — behavior that will matter enormously once these systems are operating around people rather than in labs.