Google DeepMind split the robot brain in two. Here is what each half does.

Google DeepMind released two robot models on 30 July 2026, and the interesting design choice is that they are two models rather than one. Gemini Robotics ER 2 does the thinking. Gemini Robotics 2 does the moving. This piece walks through what each layer is for, why splitting them matters, and where the public description leaves the mechanism underspecified.¹²

The problem the split solves. A robot that folds laundry needs two very different kinds of intelligence. It needs to understand the scene ("that is a shirt, it is inside-out, the sleeve is tangled") and plan ("straighten the sleeve, then fold along the shoulder seam"). And it needs to convert those plans into thousands of small motor commands per second, in a tight loop with what its cameras and joint sensors are telling it. Historically, robot systems bundled these together in a single model, or wired reasoning to hand-scripted motion primitives. That works until you change the robot, change the task, or hit anything the script did not anticipate.

DeepMind's release separates the two concerns explicitly. ER 2, the "ER" is for embodied reasoning, is described by the team as a high-level brain: it chats with humans, understands the physical world, plans multi-step tasks, and then hands off motor execution to a lower-level vision-language-action model.² Gemini Robotics 2 is that lower-level model. DeepMind describes it as their most advanced vision-language-action model, one that converts vision and language input into motor control, and that can control full humanoids from feet to fingertips as well as bi-arm robots.¹

The shape of the two-tier stack. Picture it as two loops running at different speeds. The slow loop is ER 2, thinking in seconds: what is the goal, what step am I on, has something gone wrong, what comes next. The fast loop is Gemini Robotics 2, thinking in milliseconds: given the current camera frame and joint state, what should the motors do right now. The slow loop passes intent down; the fast loop reports what actually happened back up.

The interface between them is where the design gets clever. Developers expose the low-level control interfaces, the VLA model, navigation APIs, manipulator APIs, as tools that ER 2 can call, and stream video, audio, or text directly into ER 2, according to reporting on the release by Technobezz.³ If you have followed software agents over the last two years, that phrasing should ring a bell. It is exactly the tool-calling pattern that ChatGPT, Claude and their cousins use to hit a search API or run code. DeepMind has taken the pattern and pointed it at a robot's legs.

A worked example. Technobezz reports that DeepMind demonstrated ER 2 driving a Boston Dynamics Spot: the model called navigation and manipulator APIs so the robot fetched objects on natural-language request.³ So the flow, roughly: you say "bring me the red mug". ER 2 parses that, decides it needs to (a) locate the mug, (b) navigate to it, (c) pick it up, (d) come back. For each of those, it calls a tool, the navigation API, the manipulation VLA, the way a software agent would call a database. Spot's own controllers handle the actual walking. ER 2 never has to know how a quadruped balances; it just needs to know that navigate_to(location) exists.

Why "think while moving" is the interesting claim. Buried in DeepMind's write-up is a design detail I want to flag, because it is the least obvious and the most technically loaded. ER 2, they say, is designed so the robot can think about what comes next while simultaneously performing its actions.² The pitch is: no more stop-and-think pauses between reasoning and motion.

If you know how language models work, you know why this is non-trivial. Standard LLM inference is turn-based: the model reads its input, runs a forward pass, produces output, then waits for the next input. It does not naturally produce a decision while still deliberating. Whatever ER 2 is doing to get around that — streaming output, running the reasoning model in parallel with the action model, buffering the next plan while the current one executes — the public description does not say. Call it the honest open edge of the release: we can see the capability being claimed, we cannot see the mechanism.

The numbers, and what they actually measure. Technobezz reports two benchmark figures from DeepMind's own evaluations. On moment-finding — timing precision across video-based task transitions, such as knowing when to stop pouring — ER 2 hit 91.3% accuracy with a mean absolute distance of 0.96 seconds. On progress classification across five completion bands, roughly, "how far along am I in this task, on a 5-point scale", it hit 57.4%.³

The second number is the one worth sitting with. 57.4% across five bands means the model correctly identifies its own progress state a little over half the time. That is well above chance (which would be 20% across five bands: 1 divided by 5), but a robot that misjudges where it is in a task nearly half the time will retry work that is already done, or move on from work that is not. For a system whose pitch is real-time progress tracking and mid-run adjustment, that is a specific reliability question, and one no third-party has independently checked yet.

The strategic shape. Figure AI and Tesla have pursued vertically integrated humanoids — one company, one robot, one stack. Google's release positions Gemini as an intelligence layer that OEMs plug into: the multi-robot demo reportedly featured Apptronik's Apollo 2 humanoid alongside a Franka F3 Duo dual-arm system, and Spot appeared separately as an ER 2 orchestration target.³ Whichever bet pays off, the two-tier split makes Google's bet legible: sell the brain, let others build the bodies.

What is not yet answered. DeepMind says Gemini Robotics 2 can adapt to entirely new robotic bodies in just a few hours.¹ The blog does not describe the mechanism — fine-tuning, few-shot demonstration, prompt-based adaptation. If the "few hours" claim survives contact with robots outside DeepMind's partner network, the transfer problem that has slowed robotics for a long time gets meaningfully smaller. If it does not, the two-tier architecture remains an interesting pattern, but the OEM-plug-in strategy will need a different story to tell.

Footnotes and links

Further reading

• Google DeepMind blog: https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots
• Google blog on ER 2: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2