Google DeepMind debuts Gemini Robotics 2 for whole-body humanoid control

Gemini Robotics 2 is DeepMind's push to give humanoids 'whole body intelligence' — coordinating movement across the entire physical structure rather than the upper-body-only control of the 2025 model. That means walking, crouching, and dexterous manipulation orchestrated together, enabling multi-step tasks that weren't previously possible. CEO Demis Hassabis highlighted robots reasoning through 'every movement' to tie delicate knots and teaming up to solve complex workflows.

The launch pairs the action model with Gemini Robotics ER 2, an embodied-reasoning variant that DeepMind says marks a step change in video understanding, tool orchestration, and coordination across multiple robots. Crucially, DeepMind bundled new safety measures for humanoid deployment, acknowledging the risk profile of autonomous physical AI moving from labs toward the real world. Separately, DeepMind reshuffled its Nobel-winning AI team to pursue Gemini-powered systems that assist and automate scientific research.

Competitively, this lands the same week Mistral entered Physical AI with Robostral Navigate, underscoring that robotics foundation models are becoming a contested frontier alongside language models. Google's advantage is coupling Gemini's multimodal reasoning with whole-body control; the open question is real-world robustness and safety at deployment scale. Skeptics greeted the whole-body autonomy with both excitement and scrutiny over physical AI acting in the world. What to watch: which robot hardware partners ship on Gemini Robotics 2, and whether DeepMind releases benchmarks quantifying the dexterity gains.