Google DeepMind has released Gemini Robotics 2, a suite of AI models designed to take robots from scripted motion to genuine real-time intelligence — headlined by an embodied-reasoning model that plans while it moves.

The flagship model, Gemini Robotics ER 2, acts as a high-level brain for robots: it watches what the robot perceives and simultaneously plans the next action, so the machine can 'think' about what comes next while performing its current task rather than pausing to reason between steps. DeepMind says this makes possible fluid whole-body coordination — balance, walking, and multi-step manipulation — and, for the first time at this level, multi-robot collaboration, where several robots share a reasoning process to divide and complete a task.

The suite, announced July 30 and available through the Gemini API and Google AI Studio, marks the next step in DeepMind's bet that general intelligence will be demonstrated first in the physical world. By combining video understanding with embodied reasoning, ER 2 lets a robot interpret what it sees, decide what matters, and act accordingly — closing the loop that earlier models left open between perception and action.

For robotics developers, the release lowers the barrier to building adaptive systems: no more hand-coding every contingency. The bigger question is how quickly the thinking-and-acting loop scales to real-world deployments, where latency, safety and hardware limits still dominate.

Sources