Gemini Robotics 2 brings whole body intelligence to robots

Google DeepMind has introduced Gemini Robotics 2, marking a major step toward general-purpose physical intelligence. Unlike previous generations that focused on narrow, repetitive tasks, this new system provides robots with whole-body control, advanced dexterity, and multi-robot collaboration. The update allows humanoid robots to perform complex tasks like maneuvering around cluttered spaces, picking up specific items, and coordinating with other machines to finish work faster.

The system consists of three distinct models designed for different applications. The primary Gemini Robotics 2 model converts vision and language into motor commands, allowing for control over both humanoids and bi-arm robots. Gemini Robotics ER 2 functions as a reasoning engine, enabling robots to process long-term instructions, communicate with human operators, and plan multi-step operations over several minutes. Finally, Gemini Robotics On-Device 2 provides an efficient solution for hardware that requires local processing without constant internet connectivity.

Safety remains a central focus for this release. The team introduced a new benchmark called ASIMOV-Agentic, which measures how well a robot can assess its own uncertainty and request human intervention when a task might be unsafe. The models also feature improved proximity detection, allowing robots to halt operations if people move too close. This development is part of a broader shift toward creating robots that move beyond pre-programmed routines and adapt to unpredictable real-world environments.

By enabling robots to generalize across different body types with minimal data, these models represent a shift toward physical AI that works alongside people. These advancements in dexterity and embodied reasoning bring robots closer to being functional tools in homes and workplaces. The technology is currently available through various Google platforms, with early access options for development partners looking to integrate these capabilities into their specific hardware.