Google splits the robot brain from the robot body

A robot used to ship as a single artifact: motors, sensors, and the software that turns one into the other, all from one vendor. On July 30, 2026, Google DeepMind announced Gemini Robotics ER 2, a model that breaks that package in two. ER 2 is the brain: a reasoning layer that plans the next step, calls out to Google Search or a developer-defined tool when it needs more information, and decides when to hand a subtask to a second robot. The motors and arms stay on a separate, lower-level model that the lab has been shipping for two years. A robot can now be two products from two companies, stitched together over an API.

The lower-level layer, called a Vision-Language-Action (VLA) model, is the part that actually drives the wheels and arms. It reads a camera frame, picks a joint angle, repeats. ER 2 sits above it, per the DeepMind model page, as an "embodied reasoning" orchestrator: it reasons about the task, breaks it into steps, watches the VLA execute, and steps in when the plan needs to change. The two talk through a bidirectional, streaming endpoint based on the Gemini Live API, the same kind of low-latency audio-video-text pipe that powers real-time voice agents, repurposed for robots.

The public demo runs on hardware ER 2 did not come with. Google and Boston Dynamics showed ER 2 driving a Spot quadruped through Spot's own APIs. The dog handles navigation and a manipulator arm. ER 2 sends it natural-language fetch commands. Third-party hardware, first-party planning, the body's own motor code still in the middle.

Three additions matter more than the demo. First, ER 2 can watch a task in progress. The model keeps a running read on the robot's camera and progress markers, notices when a step has failed, and self-corrects without a human in the loop. Ars Technica's coverage describes this as continuous video understanding paired with progress tracking, with the model reasoning about upcoming actions while the robot is still executing the current step.

Second, tool calling. ER 2 can pause, hit Google Search, and resume. A robot sorting recycling no longer has to rely on what was in its training set to tell a plastic bottle from a compostable cup; it can look. The same hook accepts developer-defined functions, so a warehouse integrator can wire it into an inventory system without retraining the motor model. SiliconANGLE flags tool orchestration as the headline delta from ER 1.6, and Google reports improved scores on tool-orchestration evals across real-VLA, simulated-VLA, and human tele-operation control modes, per the DeepMind model page.

Third, multi-robot collaboration in a shared space. Two robots running ER 2 can split a job, with one fetching and the other opening the door, and coordinate without a central scheduler. The feature is what makes a fleet more than a pile of Spot-shaped parts, and it is also the part of the demo most likely to meet a real-world floor covered in people who do not move on a schedule.

The split has a commercial side Google did not emphasize. When the brain and the body are separate products, three businesses can sell into the same robot: the model provider (Google), the hardware vendor (Boston Dynamics, Apptronik, the field of humanoid startups), and the integrator who wires them into a warehouse. The model provider has the most leverage, because the brain is the part customers cannot easily swap. ER 2 is the lab's bid to be the default planning service underneath a generation of robots it does not build.

The caveats sit next to the architecture. The benchmark gains Google reports are on its own tool-orchestration evals, run on its own control modes. A Spot walking a clean lab floor is not a Spot walking a hospital corridor at 3 a.m., and the demo does not show unsupervised operation in either. Tool calling with Google Search also means the model's behavior at retrieval time follows whatever the search result says, a different reliability profile from a closed system. The split shipped today. The integration is the part still unproven.