Google DeepMind releases Gemini Robotics 2 with whole-body AI control for humanoids

Google DeepMind unveiled Gemini Robotics 2, a comprehensive suite of AI models engineered to grant robots full-body autonomy, sophisticated dexterity, and the ability to collaborate across multiple machines. The system is being demonstrated on Apptronik's Apollo 2 humanoid robot, which can now execute full-body autonomous movements—walking, crouching, bending, and manipulating objects—while simultaneously reasoning through complex tasks in real time.

The release comprises three distinct models. Gemini Robotics 2 itself is a vision-language-action (VLA) model that translates visual and language inputs into robot movements, controlling both full humanoid robots and dual-arm systems. Gemini Robotics ER 2 is an embodied reasoning model capable of planning multi-step tasks, communicating with people, and enabling multiple robots to collaborate on workflows. Gemini Robotics On-Device 2 is an optimized VLA model that runs directly on robot hardware without requiring cloud connectivity, and can be adapted to new robot embodiments with only a few hours of training data.

A key advancement in Gemini Robotics 2 is whole-body control. Previous versions focused primarily on upper-body manipulation; the new model enables humanoids to coordinate movements from "feet to fingertips." In a demonstration, Apollo 2 receives the instruction "put the watering can into the green bin in the bottom shelf," then walks across the room, picks up the object, and places it in the correct location. The system also demonstrates improved dexterity, controlling Apptronik's Apollo 2 equipped with five-fingered SharpaWave hands to perform tasks such as tying a trash bag or unscrewing a light bulb. For industrial settings, the model supports conventional parallel grippers for precision insertion and packing.

Multi-robot collaboration represents another new capability. Gemini Robotics ER 2 allows different robot types to communicate and coordinate on complex workflows that would otherwise require a single machine to perform every task. The reasoning model is designed to manage task execution over several minutes, monitoring progress, recovering from failures, and adapting to changing situations. Google DeepMind describes the overall platform as the "intelligence layer" for robots, enabling machines to move beyond pre-programmed tasks and adapt to changing environments, and says it can be adapted to new hardware in a matter of hours.

Safety is integral to the release. Google DeepMind introduced ASIMOV-Agentic, a benchmark designed to evaluate "agentic safety orchestration and uncertainty resolution," measuring a robot's ability to refuse unsafe actions, determine when a task cannot be completed safely, and request human intervention when appropriate. Gemini Robotics ER 2 also improves human-awareness capabilities, enabling robots to detect nearby people, trigger safety functions, and stop safely when someone enters their working area. The announcement builds on the companies' ongoing collaboration; earlier this month, Apptronik launched Robot Park, a facility where Apollo 2 robots collect data used to train next-generation AI models including Gemini Robotics 2. Gemini Robotics ER 2 is now available through Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while Gemini Robotics 2 and Gemini Robotics On-Device 2 are available to selected early-access partners.