Google DeepMind Launches Gemini Robotics 2, Giving Humanoid Robots Whole-Body Control
Google DeepMind unveiled a three-model Gemini Robotics 2 suite that lets humanoid robots coordinate torso, arms and legs together instead of controlling only the upper body, demonstrated on Apptronik's Apollo 2 robot with a 92% success rate on a full-body dexterity task.
Google DeepMind announced Gemini Robotics 2 on July 30, moving past the prior generation's upper-body-only control to let robots plan and adjust movement from the torso down to the legs in real time. The suite is being demonstrated on Apptronik's Apollo 2 humanoid, which can now walk, crouch, bend and manipulate objects while reasoning through multi-step tasks autonomously.
- Gemini Robotics 2: vision-language-action model that converts visual and language input into full-body robot movement
- Gemini Robotics ER 2: embodied-reasoning model for multi-step task planning and coordinating multiple robots at once
- Gemini Robotics On-Device 2: a compressed VLA variant that runs locally on robot hardware without a cloud connection
- Apollo 2 hit a 92% success rate unscrewing a light bulb, a task DeepMind uses as a benchmark for precise full-body coordination
The release extends Google's model lineup beyond text and code into embodied AI, following Mistral's Robostral Navigate into the robotics space, though Gemini Robotics 2 targets full-body humanoid control rather than single-camera navigation.