Overview
- Google DeepMind publicly announced Gemini Robotics 2 on July 30–31, 2026 as a three-part robotics intelligence stack that links a vision-language-action controller, an embodied-reasoning planner (ER 2), and an on-device action model.
- DeepMind published ER 2 performance numbers showing 57.4% accuracy on continuous progress classification, 91.3% accuracy on moment finding, and a mean timing error of 0.96 seconds to measure the model’s temporal reasoning.
- Video demos showed the system controlling Apptronik’s Apollo 2 humanoid, Boston Dynamics’ Spot and Franka platforms to walk, grasp with five-finger hands, tie knots, seal bags, and coordinate handoffs between robots.
- Gemini Robotics ER 2 is available to developers via Google AI Studio and private preview on the Gemini Enterprise Agent Platform, while the VLA and On-Device models are in early-access partner programs with fast local adaptation touted at under 200 examples.
- Google introduced a new safety benchmark called ASIMOV-Agentic and layered human-proximity protections that can halt robots, but the company cautioned that speed and some task-accuracy limits remain and that broader deployment is still at an early, partner-focused stage.