Gemini Robotics, explained: embodied reasoning in a VLA

Ask a Gemini Robotics arm to clear a cluttered table into the right bins and something unusual happens before it moves. The model writes a short plan in plain English: the banana is a fruit, so it goes to compost; the crumpled receipt is paper, so it goes to recycling; the phone is neither, so leave it alone. Only then do the grippers start. That legible pause, a robot reasoning out loud before it touches anything, is the clearest tell of what Google DeepMind has been building toward.

The surprising design choice is that Gemini Robotics is not one model. DeepMind shipped a pair: a vision-language-action model that emits motor commands, and a separate reasoning model, Gemini Robotics-ER, where ER stands for embodied reasoning. The split is the whole argument. The hard part of a general-purpose robot policy is not producing joint angles; it is knowing where the cup is, which edge to grasp, what happens if you push the stack, and in what order to do things. DeepMind's bet is that this world understanding is worth factoring out into its own component you can pretrain once and reuse.

This piece walks through how that fold works: what embodied reasoning adds on top of a plain multimodal model, how the reasoning half and the acting half talk to each other, and where the design pays off and where it strains. It is written for people who already know what a VLA is and want the architecture, not the launch video.

Two models, one robot

Gemini Robotics began as an action extension of Gemini 2.0. The base is a large multimodal model that already reads images and text; DeepMind added physical action as an output modality, so the same network that can describe a scene can also propose how to move through it. That is the VLA. On its own it leans heavily on real teleoperated demonstrations, with the ALOHA 2 bimanual platform as the primary body it learned on.

Gemini Robotics-ER is the other half, a vision-language model tuned until its picture of the world is sharp enough to drive a robot: 2D and 3D object detection, pointing at a named part with a pixel coordinate, predicting a grasp, sketching a trajectory, and matching the same object across several camera views. These are not abstract talents. Each maps to a question a controller must answer before it can act. A plain chat model can tell you a mug has a handle; ER can tell you which pixels the handle occupies and where a gripper should close.

What embodied reasoning buys you

The reason DeepMind did not simply aim Gemini 2.0 at a gripper is that language pretraining teaches a model to talk about the world, not to locate things in it with the precision a motor needs. Embodied reasoning is the layer that closes that gap. It is spatial (where, how far, what orientation), physical (what is rigid, what will tip, what sits behind what), and temporal (what has to happen before what else).

Because this layer is a model rather than a hand-tuned perception stack, it inherits the base model's generality. You can ask it about objects and arrangements it never saw in any robotics dataset, and it will often still point, still estimate, still order the steps, because the underlying Gemini has met the concept somewhere. That is the payoff of building reasoning on a foundation model instead of bolting a detector onto a policy: the long tail of odd objects and odd instructions is handled by transfer, not by more labels.

The wager is that embodied reasoning is a capability you pretrain once and reuse across robots, not a skill you grind out one machine and one kitchen at a time.

Thinking before acting

The 1.5 generation made the fold explicit. Gemini Robotics 1.5 produces an internal chain of natural-language reasoning before it commits to any action, the visible thinking that opens this article. Its partner, Gemini Robotics-ER 1.5, is promoted to an orchestrator: it takes a high-level goal, breaks it into steps, can call digital tools such as web search to fill a knowledge gap, and hands each concrete subtask to the action model to execute. One model plans and reasons; the other moves.

Two consequences matter. First, the reasoning is legible, which makes failures easier to diagnose: you can read the plan and see whether the mistake lived in the thinking or in the hands. Second, skills learned on one body can move to another. DeepMind describes carrying competence across embodiments, from the ALOHA 2 bimanual rig to bi-arm Franka arms and the Apptronik Apollo humanoid, without retraining each from scratch. Cross-embodiment is the same problem Open X-Embodiment framed for the whole field: pool experience across different robots so a policy is not trapped in one chassis.

The family, and how it compares

By late 2025 the lineup had four notable members, and reading them together shows the trajectory from cloud-bound demo toward something you could ship.

Set against its peers, the design reads as a deliberate contrast. NVIDIA's Isaac GR00T leans on simulation to manufacture training data at volume. Physical Intelligence pools real teleoperation across many robot types into a single policy. DeepMind's distinctive move is orthogonal to both: keep the data real, but factor reasoning out of action so the expensive-to-learn world model is shared, and let a smaller action head specialize per body.

Where it strains

A two-stage brain-and-body system buys clarity and pays in latency and coordination. A planner that reasons in text, sometimes calls a tool, then dispatches to an action model is slower than a single reflex policy, which is precisely why an on-device VLA exists: part of the loop has to run locally at motor rates, without a round trip to the cloud. There is a division-of-labor question underneath, too, deciding how much to trust the plan versus the reflexes when the two disagree.

Generalization has edges as well. Embodied reasoning transfers impressively across objects, but a reasoning model can still be confidently wrong about physics it has only read about, and a legible plan that is wrong is still wrong. DeepMind pairs the releases with safety work, including the ASIMOV benchmark for judging whether a model's decisions are physically and semantically safe, a quiet admission that a model fluent enough to plan is also fluent enough to plan something it should not.

What to watch

The interesting question is not whether Gemini Robotics can fold an origami fox on a demo stage. It is whether the reasoning-and-action split holds up as tasks get longer and robots get cheaper. If embodied reasoning really is a reusable layer, the same brain should light up a warehouse arm and a home humanoid with mostly a fresh action head between them. If it is not, the field slides quietly back toward one model per robot, per room. Watch the transfer results, not the highlight reels.