(WS News) – Google DeepMind has released Gemini Robotics 2, a new suite of three AI models designed to give robots more advanced physical reasoning and control, marking one of the most significant pushes yet by a major AI lab into embodied, real-world robotics.
The suite is built around three distinct components. The first is a vision-language-action model aimed at whole-body humanoid control, allowing robots to interpret their surroundings and translate that understanding directly into coordinated physical movement. The second, an embodied-reasoning model referred to internally as ER 2, focuses on multi-step planning and coordination between multiple robots working on the same task. The third is a lightweight, on-device variant built to adapt quickly to new robot bodies, with DeepMind saying it can adjust to unfamiliar hardware configurations within hours rather than the weeks such calibration has traditionally required.
Together, the models represent an attempt to solve one of robotics’ longest-standing bottlenecks: building AI systems flexible enough to generalize across different robot bodies and tasks, rather than requiring bespoke training for every new machine. The multi-robot collaboration capability in particular points toward warehouse, logistics and manufacturing environments where fleets of robots need to coordinate rather than operate in isolation.
The release comes amid a broader surge of investment and activity in humanoid and physical AI. Governments and private investors alike have been pouring money into robotics infrastructure this year, even as some jurisdictions have moved to restrict or scrutinize the import of foreign-made humanoid robots on national security grounds. Washington, for instance, has separately been investing in faster chip-to-chip connections for AI clusters while placing new restrictions on Chinese humanoid robots.
For DeepMind, Gemini Robotics 2 builds on the company’s earlier robotics efforts and reflects a wider industry shift away from purely text- and image-based AI models toward systems that can act directly in the physical world. Rivals across the AI landscape have been racing to make similar leaps, with several major labs and hardware makers unveiling their own humanoid and agentic robotics platforms in recent months.
DeepMind has not detailed a full commercial rollout timeline for the new suite, but the announcement signals that the next phase of competition in AI is increasingly playing out not just in chatbots and software agents, but in machines that can see, plan and physically act in the real world.
