Google's artificial intelligence is breaking free from the screen to take total control of robotic motility. With the release of Gemini Robotics 2, DeepMind has evolved its strategy: moving beyond simple upper-body commands for tabletop tasks to a coordinated management of every single joint, from feet to fingertips.

Add AlexTech.ai asPreferred Source on Google

One Brain for Every Robotic Body

The core innovation lies in a three-tier architecture. The Gemini Robotics 2 model acts as the vision-language-action interface, translating sensory inputs into motor commands. Above it sits Gemini Robotics ER 2, the reasoning layer that plans multi-step jobs, utilizes Google Search, and coordinates heterogeneous teams—such as a wheeled machine and a humanoid splitting a task.

Finally, the On-Device 2 version allows for local execution without internet connectivity. Crucially for the industry, it can adapt to an entirely new robot body with fewer than 200 examples and only a few hours of training, solving one of robotics' most persistent hurdles.

The Dexterity Gap and Physical Limits

Despite impressive demos with Apptronik's Apollo 2, the data reveals a gap between vision and reality. While the AI can unscrew a light bulb 92% of the time, precision plummets for finer tasks: sealing a ziplock bag or tying a trash bag shows success rates between 40% and 44%. Robots also remain slow, pausing to think through movements that humans perform instinctively.

Safety and Hardware Geopolitics

The expansion of "Physical AI" brings concrete risks. To mitigate this, Google introduced ASIMOV-Agentic, a benchmark to test if the reasoning model can refuse unsafe commands or flag impossible tasks. Simultaneously, Google faces a hardware dilemma: while partnering with Western firms like Apptronik and Boston Dynamics, US bans on Chinese-made robots complicate access to many of the physical platforms these models are designed for.

The Path to Physical AGI

The ultimate goal is Physical AGI—an artificial general intelligence capable of performing any human physical task. This release marks the transition from theory to practice, placing Google in direct competition with OpenAI and Nvidia in the race for a universal model that can power any robotic body.