Google DeepMind Unveils Gemini Robotics 2 for Whole-Body Humanoid Control
A new tiered AI architecture separates high-level reasoning from motor execution to enable complex, multi-step robotic tasks.
Google DeepMind launched Gemini Robotics 2 and Gemini Robotics ER 2 on July 30, 2026, marking a significant leap toward general-purpose physical AI. The system introduces whole-body intelligence for humanoid robots, allowing them to coordinate legs, torso, and hands as a single integrated unit.
The new architecture operates on a tiered system. Gemini Robotics 2, a Vision-Language-Action (VLA) model, manages low-level motor execution, enabling robots to walk, crouch, and bend fluidly. Above this sits Gemini Robotics ER 2, an embodied reasoning model that functions as a high-level "brain." This reasoning layer allows robots to interact with humans, understand their physical surroundings, and plan multi-step tasks, including the ability to orchestrate other APIs or call tools like Google Search.
The Shift to Integrated Control
Historically, robotic AI often treated limbs as separate control systems or remained confined to static tabletop environments. Gemini Robotics 2 moves away from this fragmented approach by integrating vision, language, and action into a unified framework. This allows the AI to be deployed across diverse hardware forms, ranging from wheeled rovers to full humanoids, treating the entire physical form as one cohesive system rather than a collection of independent parts.
Implications for Physical AI
This transition represents a shift from simple task execution to "agentic" physical AI. By pairing high-level reasoning with low-level execution, robots can now self-correct in real-time using video feedback and collaborate with other machines through a shared semantic understanding.
The system also demonstrates advanced "temporal intelligence," achieving 91.3% accuracy in precision moment-finding—the ability to identify exact video frames for task transitions—and 57.4% accuracy in continuous progress classification. These capabilities allow for safer and more precise operation around humans through improved proximity detection.
Industry Adoption and Next Steps
Google DeepMind has already established a network of launch partners to deploy the technology. The system has been demonstrated with Boston Dynamics' Spot, as well as Apptronik's Apollo 2 and hardware from Agile Robots. As these models move from controlled demonstrations to real-world environments, the industry will be watching how effectively the ER 2 reasoning layer handles unpredictable human environments and whether the VLA model can maintain its precision across different robotic morphologies.