Google DeepMind Unveils Gemini Robotics 2 Model

Google DeepMind introduced Gemini Robotics 2, a vision-language-action model enabling whole-body coordination for humanoid robots from partners like Boston Dynamics and Apptronik.

Benchmark success rates for whole-body tasks range from 45.7% to 76.3%, with multi-finger dexterity scores varying between 32% and 92% depending on the specific task complexity.

The system includes safety features to detect and halt motion near humans, though access is currently restricted to a 'Trusted Tester' waitlist for developers.