Google DeepMind releases Gemini Robotics 2, enabling full-body humanoid control with 22-DoF hands

Google DeepMind/Gemini Robotics

Google DeepMind unveiled Gemini Robotics 2, its latest vision-language-action model for robotic control. The system features three components: VLA 2 for physical action, Embodied Reasoning 2 for high-level planning, and On-Device 2 for local operation. It can control whole-body humanoid movement including walking, crouching and complex manipulation with five-fingered hands featuring 22 degrees of freedom.

• Full-body control: walking, crouching, reaching, and complex manipulation.
• 22 degrees of freedom in five-fingered hands for dexterous grasping.
• Embodied Reasoning 2 as high-level planning brain for multi-step tasks.
• On-Device 2 runs locally without network connectivity.
• Adapts to new robot embodiments with just hours of training data.