Google DeepMind Ships Three Physical AI Models For Whole Body Control, Dexterity And Multi Robot Collaboration
What happened
Google DeepMind launched Gemini Robotics 2, a new AI intelligence layer designed to control physical robots. This release delivers three distinct AI models focused on whole body humanoid control, embodied reasoning for task management, and fast on-device adaptation to different robot bodies. One combined checkpoint can operate the Apptronik Apollo 2 humanoid and the Franka Duo robotic arm. However, only the Gemini Robotics ER 2 model, aimed at task orchestration and embodied reasoning, is currently available to the public.
Why it matters
Controlling robots in the real world with complex, coordinated movement and reasoning has long been a bottleneck for robotics applications. Gemini Robotics 2 targets this gap with specialized AI models that combine vision, language, and action. The vision-language-action model facilitates detailed whole-body humanoid control, pushing robotics beyond basic scripted behaviors. Meanwhile, ER 2 enables multi-step task orchestration in real environments where robots need to reason about objects and goals. Fast on-device adaptation cuts the time and effort needed to deploy these models across various robot platforms. For users, that translates to more capable, flexible robots with fewer integration headaches.
The integrated checkpoint driving both a humanoid robot and an industrial robot arm signals a step toward unifying control across different hardware types. While the full suite is not publicly available yet, releasing ER 2 publicly offers developers a resource to build on for embodied AI experimentation. The release challenges robotic system builders to think about AI as a modular intelligence layer that can generalize across robot bodies, simplifying deployment and accelerating development cycles.
What to watch next
Watch for how quickly the broader robotics community adopts ER 2 and what projects emerge using it. Its effectiveness on diverse robots outside DeepMind’s ecosystem will reveal practical utility and limitations. Also monitor when and if Google DeepMind releases the other two models publicly, as that would expand access to whole body control and on-device adaptation capabilities. Finally, observe what partners like Apptronik do with this technology commercially. Their success or struggles will provide early signals on how quickly physical AI models translate into new industrial or service robot deployments.
AI Quick Briefs Editorial Desk
