Gemini Robotics ER 2 on Google Cloud: the "high-level brain" that adapts to any robot

Gemini Robotics ER 2 lands on Google Cloud: the universal brain for robots is (finally) a product

🔎 A model that thinks, no matter what body you give it

On October 27, 2026, Google crossed a milestone the robotics industry had been waiting years for: the general availability of Gemini Robotics Embodied Reasoning 2 (ER 2) on Google Cloud. It's not a robot. It's not even a model that directly controls motors. It's something more interesting: a high-level decision-making brain, designed to adapt to any robotic body — fixed industrial arm, mobile manipulator, or humanoid.

Why does this matter now? Because robotics has always run up against an economic wall: every robot, every task, every environment requires years of specialized engineering. A model that reasons above the hardware, and plugs into low-level action models, breaks that equation. Google DeepMind laid it all out in its official announcement, and making it available through Google Cloud turns a lab feat into an industrial building block.

The question is no longer "can we give a robot general intelligence?" but "which vendor sells the best robotic brain?". Nvidia has its answer with Isaac GR00T. Google has just provided a second one, radically different in its architecture.

The essentials

• Gemini Robotics ER 2 is available in GA on Google Cloud as of October 27, 2026, after a preview phase with industrial partners.
• It's a high-level "embodied" reasoning model: it understands spatial and temporal video, plans over several minutes, and breaks down objectives into hundreds of successive decisions.
• It is body-agnostic: it orchestrates low-level action models (VLAs or specialized policies) rather than directly piloting the actuators, making it compatible with fixed arms as well as humanoids.
• Safety guardrails are natively integrated, with action limits verified before execution.
• It supports multi-robot collaboration, with several machines sharing the same reasoning to coordinate their tasks.
• It positions itself above frameworks like Nvidia Isaac GR00T, which focus more on low-level control policies.

Recommended tools

What is Gemini Robotics ER 2, concretely?

ER 2 is a reasoning model specialized for robotics, not an action model. The distinction is fundamental and sums up Google DeepMind's entire strategy.

The first generation of Gemini Robotics, launched in 2025, was a Vision-Language-Action (VLA) model: it received images and directly produced motor commands. Powerful, but rigid — the model had to be trained or fine-tuned for each type of body. Google then introduced Gemini Robotics-ER 1.5 in September 2025, which already separated reasoning from control. ER 2 brings this separation to maturity and makes it commercially available.

Concretely, ER 2 occupies the cognitive layer:

• Perception: it ingests video streams (RGB, depth) and builds a spatial AND temporal representation of the scene — it understands not only where objects are, but what just happened and what is currently moving.
• Planning: faced with a goal ("prepare this order", "tidy up this workshop"), it breaks the task down into a sequence of hundreds of decisions over several minutes, handling unexpected events along the way.
• Orchestration: it sends sub-goals to low-level action models — its own Gemini Robotics VLAs, or third-party policies — which translate intentions into actual movements.

This is exactly the same logic that structures modern software agents: a powerful reasoning model at the top, specialized executors below. Google is applying to the physical world what already works in the digital world — and if you follow the evolution of Google's models, from Gemini 3.1 Pro to ambient agents like Gemini Spark, the coherence of the strategy is plain to see: a reasoning core, deployed across every domain.

Why does separating the brain from the body change everything?

Because it's the only economically viable path to generic robots at scale. A monolithic VLA must be retrained for every morphology; a high-level brain adapts.

Think about what that implies. Today, an integrator deploying a robotic arm in a factory spends months on task-specific programming: every gripper, every workshop layout, every part variation requires custom work. With ER 2, the reasoning is shared — it's the same model that understands a laboratory scene, a kitchen, or an assembly line. Only the execution layer changes.

This layered architecture resembles what computing did with operating systems:

1. The hardware layer: motors, grippers, sensors — specific to each robot.
2. The action layer: low-level policies that know how to "grasp," "insert," "fold" — often compact models fine-tuned per body.
3. The reasoning layer: ER 2 — scene understanding, planning, safety, coordination.

Google isn't the only one believing in this stack. Nvidia is pushing the same idea with its Isaac stack, but with a different philosophy: GR00T is an open weights model focused on control policies, designed for humanoids. Google, for its part, sells the top of the stack and remains agnostic about the rest. Two strategies, one same conclusion: value is shifting toward reasoning.

My take? This is the right battle. A universal brain is more defensible than a grasping policy, because it benefits from all the advances of frontier models — ER 2 directly inherits the progress of the Gemini lineage, whose generalist models like Gemini 3.1 Pro or OpenAI's GPT-5.5 show every quarter just how fast multimodal reasoning is improving.

What ER 2 Can Do: Spatial Video, Long-Horizon Planning, Adaptation

ER 2 combines three capabilities that, chained together, enable a robot to handle tasks lasting several minutes without human supervision.

Spatial and Temporal Video Understanding

Most robotic systems reason over isolated images. ER 2 reasons over a continuous video stream: it tracks a moving object, anticipates a trajectory, understands that a door has just been opened or that a person has just entered its field of action. This temporal dimension is what is sorely missing from systems that analyze isolated frames — and it's what makes it possible to react to a dynamic world rather than to snapshots.

Planning Over Several Minutes

Faced with a high-level goal, ER 2 doesn't produce a single action: it produces a plan broken down into hundreds of successive decisions. Take a dish out of the freezer, check its temperature, pick the right oven, monitor the cooking, alert if something goes wrong. Each step is continuously re-evaluated, and the plan repairs itself when reality diverges — an object has been moved, a tool is missing, a human intervenes.

Orchestration of Action Models

ER 2 doesn't reinvent the wheel. It delegates execution to specialized action models, whether proprietary (Google's VLAs) or third-party. It's an orchestrator role: the model reasons, checks feasibility, sends sub-goals, and monitors the outcome. If an action model fails, ER 2 detects it and replans.

Multi-Robot Collaboration

A final piece, often underestimated: ER 2 can coordinate multiple robots that share the same reasoning. Two arms splitting an assembly task, a mobile robot resupplying fixed stations, a humanoid taking over from a fixed arm in an area it can't reach. Multi-agent coordination is hard even in pure software; making it work in the physical world, with real latencies and bodies that get in each other's way, is a genuine differentiator.

Safety: Built-in guardrails, not bolted on after the fact

ER 2 natively integrates safety guardrails that verify actions before they are executed. This isn't a marketing detail — it's probably the decisive argument for industrial players.

A robot that reasons is a robot that can be convincingly wrong. Classic safety limits (safety cells, emergency stops, virtual zones) protect against predictable mechanical behaviors, but not against the judgment errors of a generative model: grabbing the wrong object, applying inappropriate force, interpreting an ambiguous instruction in a dangerous way.

Google's approach with ER 2 is to make safety a step in the decision pipeline:

• Pre-action verification: each proposed action is evaluated against defined limits (forbidden zones, object types that must not be handled, maximum forces) before being sent to the action model.
• Layer separation: the reasoning model never directly drives the motors; low-level policies act as an additional containment layer.
• Controlled degradation: in case of uncertainty or a situation outside its area of competence, the model can stop and request human intervention rather than improvising.

This is the same debate running through software AI regarding autonomous agents — where the best-performing models on agentic tasks, like GPT-5.5 (98.2 on agentic benchmarks according to June 2025 rankings) or Claude Opus 4.7, must arbitrate between autonomy and control. In the physical world, the stakes are simply more concrete: a software error corrupts data, a robotic error injures someone. The fact that Google makes guardrails a central selling point rather than a footnote is a sign of maturity for the entire industry.

ER 2 vs Nvidia Isaac GR00T: two philosophies of the robotic brain

ER 2 and Isaac GR00T are not direct competitors: one sells reasoning, the other provides execution policies. But they sketch out two visions of the robotics stack, and integrators will have to pick a side — or combine them.

Nvidia has done considerable work to structure the humanoid ecosystem: Isaac GR00T provides fine-tunable foundation models, Isaac Sim enables large-scale simulation, and the company has established its hardware references with platforms like the reference humanoid presented at GTC Berlin. Impressive demonstrations are also plentiful on the applied research side, as illustrated by the Nvidia ASPIRE V10 robot that solves a Rubik's Cube by writing its own skills.

But Nvidia's strength — open, high-performing low-level policies — is also its limitation: someone still has to decide what to do, when, and how to coordinate multiple machines. That's precisely the gap ER 2 is here to fill.

My take: the two are complementary more than competitive. A realistic stack in 2026-2027 could be ER 2 as the brain, GR00T-style policies for execution, and hardware from a third-party integrator. Google doesn't forbid it — ER 2 is designed to orchestrate third-party action models. Nvidia, on the other hand, obviously wants its own stack to cover everything, from reasoning down to the motor. The commercial battle will be fought on this point: who captures the customer relationship at the top of the stack?

How to access ER 2 on Google Cloud

ER 2 is accessible via Vertex AI, Google Cloud's managed ML platform, generally available since October 27, 2026. No weight downloads, no self-hosting: it's a service.

The typical path for a team:

1. Provision access via Vertex AI on a Google Cloud project, with the quotas associated with the Gemini Robotics service.
2. Connect the robot's streams: cameras (RGB, depth), robot state, telemetry — ER 2 consumes video and states, not proprietary formats.
3. Plug in the action models: either Google's Gemini Robotics VLAs, or your own policies, via the orchestration interface.
4. Configure the safeguards: forbidden zones, excluded objects, force thresholds, human escalation rules.
5. Test in simulation before any physical deployment — a non-negotiable step.

To check the availability of the service from your environment, a simple CLI command is enough:

gcloud ai models list --region=us-central1 --filter="display_name:gemini-robotics"

The cloud-only choice is deliberate and consistent with Google's strategy: as with its reasoning models, the company keeps the weights and monetizes access. For industrial players, it's a trade-off to evaluate seriously — operational simplicity versus dependence on a single infrastructure. Teams that require full control of the weights will turn to open weights alternatives; teams that want to deploy quickly without a dedicated ML team will find in Google's offering a considerable shortcut.

Why now? The context of an industrial shift

Three convergences explain why ER 2 is arriving on the market in October 2026, and not two years earlier or later.

First convergence: the maturity of multimodal reasoning. The frontier models of 2025-2026 — Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7 — have crossed a threshold in video understanding and long-horizon planning. ER 2 builds directly on this lineage: without these advances, a robotic brain capable of planning over minutes was simply not feasible.

Second convergence: the massive arrival of humanoids. Platforms are multiplying — Figure, Agility, Unitree, and Nvidia's references. But a humanoid without a generic brain is just an expensive automaton. The bottleneck has shifted from hardware to intelligence, and brain providers are now competing for this market.

Third convergence: pressure from industrial players. Logistics, manufacturing, laboratories: sectors facing labor shortages no longer want proof of concept, they want deployable systems with safety guarantees. General availability on Google Cloud — with SLAs, quotas, and support — is the answer to this demand.

What strikes me is the speed at which the sector is professionalizing. Two years ago, general robotics was a research topic. Today, we compare cloud offerings, pricing, architectures — exactly the way we compare LLMs. The field has just moved from science to engineering.

❌ Common Mistakes

Mistake 1: Believing that ER 2 directly controls the motors

ER 2 is a reasoner, not a controller. It produces intentions and sub-goals, which are delegated to action models. If you plug ER 2 into a robot without an execution layer, it won't move an inch. Solution: plan for a low-level policy from the start (VLA Gemini Robotics, a fine-tuned GR00T model, or a homegrown policy).

Mistake 2: Underestimating the sensor integration work

"Compatible with any robot" does not mean "plug-and-play". ER 2 needs quality video streams, proper spatial calibration, and reliable telemetry. A poorly positioned camera degrades the entire reasoning chain. Solution: invest in perception quality before dreaming of advanced planning.

Mistake 3: Deploying to production without going through simulation

Built-in safeguards reduce risks, they don't eliminate them. Testing directly on a physical robot means exposing expensive hardware — and potentially people — to unvalidated behaviors. Solution: systematic validation in simulation, followed by progressive deployment with human supervision during the first cycles.

Mistake 4: Ignoring the cloud dependency issue

ER 2 is a cloud service: no network, no reasoning robot. In industrial environments, a loss of connectivity must be anticipated. Solution: design a degraded mode — low-level policies keep executing the tasks in progress, and the robot stops cleanly when it comes to new decisions.

❓ Frequently Asked Questions

Does ER 2 replace Google's VLA models?

No, it complements them. Gemini Robotics VLAs remain the execution layer that translates intentions into movements. ER 2 sits above them: it understands the scene, plans, delegates, and supervises. The architecture is layered, not a replacement.

Does ER 2 work with any robot?

In theory, yes: the model is embodiment-agnostic, compatible with both fixed arms and humanoids, and can orchestrate third-party action models. In practice, quality depends on the execution layer and the perception available on your hardware. Google validated industrial partnerships during the preview phase.

What is the price of ER 2 on Google Cloud?

Pricing is usage-based via Vertex AI, and Google does not publish a simple price grid comparable to that of text LLMs. The cost depends on the volume of video streams processed and the frequency of decisions. Check current pricing on cloud.google.com (October 2026).

Does ER 2 make Nvidia Isaac GR00T obsolete?

No. GR00T excels at low-level policies and remains open weights — an asset for anyone who wants control over their weights. ER 2 occupies the reasoning layer that GR00T does not cover. The two can even be combined within the same stack.

Can you self-host ER 2?

No. Google distributes ER 2 exclusively as a managed service via Google Cloud, in line with its strategy for its frontier models. Teams requiring full control over the weights will have to turn to open alternatives, with the associated reasoning limitations.

Does ER 2 really handle multi-robot collaboration?

Yes, it's a native capability announced by DeepMind: multiple robots can share the same reasoning to coordinate their tasks. This is one of its key differentiators compared to solutions focused on a single embodiment.

✅ Conclusion

With ER 2 on Google Cloud, Google isn't selling a robot: it's selling the control center of general robotics — and that's probably the most profitable position in the entire stack. The next milestone to watch: the first large-scale industrial deployments, and Nvidia's response on the reasoning layer. To understand how Google's models stack up against the competition for everyday use cases, check out our Gemini vs ChatGPT vs Claude comparison.