Gemini Robotics ER 2: Google DeepMind's Embodied-Reasoning Robot Brain, Explained

Google DeepMind's Gemini Robotics ER 2 — released in public preview on July 30, 2026 — is a specialized vision-language model built to be the high-level "brain" of a robot. "ER" stands for Embodied Reasoning: rather than directly driving motors, ER 2 sees and understands the physical world, reasons about space and time, plans multi-step tasks, orchestrates tools, and coordinates robots. It also ships a real-time streaming variant for low-latency, interactive control. This guide explains what ER 2 is, how it differs from a direct-control model, what's new versus ER 1.6, and where it fits — with capabilities sourced.

Accuracy note: capabilities, availability, and safety claims come from Google DeepMind's announcement and the Gemini API change log (2026-07-30). Robotics benchmarks are early and vendor-reported; pricing follows standard Gemini API billing and specifics weren't published. Details change — verify before building.

TL;DR. Gemini Robotics ER 2 is an embodied-reasoning VLM that acts as a robot's planner and perception brain: video understanding, spatial reasoning, multi-step tool orchestration, video moment localization, progress tracking, multi-robot coordination, and self-correction, with a streaming variant for real-time bidirectional audio-video interaction. It's available in the Gemini API (code>gemini-robotics-er-2-preview/code> and code>gemini-robotics-er-2-streaming-preview/code>) and Google AI Studio, as a closed preview. Google positions it as its safest robotics model yet, with notable gains over ER 1.6 in multi-robot coordination and progress tracking. It's a reasoning layer, not a low-level actuator controller.

Key takeaways

• Embodied Reasoning (ER): ER 2 is the high-level perception-and-planning brain, not a direct motor controller (that's the separate VLA/on-device line).

• Core skills: video understanding, spatial reasoning, multi-step tool orchestration, video moment localization, progress classification, multi-robot coordination, self-correction.

• Streaming variant: code>gemini-robotics-er-2-streaming-preview/code> adds low-latency, bidirectional audio-video interaction for real-time control and dialogue.

• Safety-first: Google positions ER 2 as its safest robotics model on safety-constraint adherence and human-proximity, with ASIMOV2 benchmark gains.

• Availability: Gemini API + Google AI Studio (public preview); Enterprise Agent Platform in private preview; VLA/on-device models limited to early partners. Closed.

What "Embodied Reasoning" means

Modern robotics stacks increasingly split into two layers. A high-level reasoning brain understands the scene, decides what to do, and plans; a low-level action model translates plans into precise motor commands. Gemini Robotics ER 2 is the first layer: an embodied-reasoning vision-language model. It takes in video (and audio and text), builds an understanding of objects, space, and progress over time, and outputs plans and tool calls — including invoking external tools like Google Search — that a separate action system or human can execute. In other words, ER 2 is the part of the robot that thinks, not the part that moves; Google's direct-control vision-language-action (VLA) and on-device models are a separate line, currently limited to early partners.

What ER 2 can do

Google describes a broad embodied-reasoning skill set. ER 2 can understand video and localize specific moments within it ("when did the cup fall?"), reason about 3D space and object relationships, classify task progress (is the step done, partway, or failed?), and orchestrate multi-step tool use to accomplish a goal. It can hold a dialogue with a person and plan tasks that span several minutes, self-correct when something goes wrong, and — notably — coordinate multiple robots working together. These are exactly the capabilities a robot needs to operate in messy, real environments rather than scripted demos.

The streaming variant: real-time interaction

Alongside the standard preview, Google shipped code>gemini-robotics-er-2-streaming-preview/code>, optimized for low-latency, bidirectional audio-video. That matters for embodied use: a robot needs to perceive, reason, and respond continuously, not in slow request-response turns. The streaming model enables real-time interaction — watching a live feed, talking with a person, and adjusting a plan on the fly — which is essential for interactive assistants, teleoperation support, and dynamic multi-step tasks.

What's new versus ER 1.6

ER 2 is a clear step up from Gemini Robotics ER 1.6. Google highlights markedly better multi-robot coordination and task-progress tracking, stronger multi-step tool orchestration, and improved safety behavior. On the safety side specifically, Google positions ER 2 as its safest robotics model to date, citing improved adherence to safety constraints and human-proximity behavior and gains on the ASIMOV2 safety benchmark. For teams building real deployments, the coordination, progress-tracking, and safety improvements are the most consequential upgrades.

Safety and evaluation

Safety is central to ER 2's positioning. Google reports that it leads on safety-constraint adherence and on how closely its behavior matches safe human judgment around people, with improved scores on ASIMOV2 (a safety-oriented robotics benchmark). Because robotics acts in the physical world, these safety properties matter far more than they do for a chatbot — a planning error can cause real harm. As with any vendor benchmark, treat the specifics as early and directional; independent robotics evaluation is still maturing.

Availability and pricing

ER 2 is available in public preview through the Gemini API — model ids code>gemini-robotics-er-2-preview/code> and code>gemini-robotics-er-2-streaming-preview/code> — and in Google AI Studio. An Enterprise Agent Platform integration is in private preview, and the direct-control VLA and on-device models remain limited to early partners. It's closed. Pricing follows standard Gemini API billing; Google didn't publish robotics-specific rates at launch. Confirm current access and costs in the Gemini API documentation.

Three scenarios where ER 2 fits

1. Warehouse and logistics robots

Spatial reasoning, progress tracking, and multi-robot coordination suit pick-and-place, sorting, and fleet tasks where robots must understand scenes and cooperate.

2. Interactive assistant robots

The streaming variant's real-time audio-video interaction enables robots that watch, listen, converse, and plan multi-minute tasks with a person.

3. Inspection and monitoring

Video understanding and moment localization make ER 2 useful for analyzing live or recorded feeds — spotting events, classifying states, and flagging progress or failures.

How it fits a broader AI stack

Gemini Robotics ER 2 is a specialized robotics model accessed through Google's Gemini API, not a general-purpose LLM — so it isn't something a general model router hosts. For the general-purpose language and multimodal parts of an application around a robot — natural-language interfaces, reasoning, orchestration, analytics — a vendor-neutral endpoint keeps your options open: a href="https://www.orcarouter.ai/">OrcaRouter/a> provides one OpenAI-compatible endpoint across many text and multimodal LLMs (including Google's general Gemini models), while ER 2's embodied control runs through Google's dedicated robotics API. That separation is natural: use the specialized robot brain for embodiment, and a vendor-neutral endpoint for everything else.

Limitations and caveats

ER 2 is a reasoning/planning layer, not a turnkey robot — you still need an action system (Google's VLA/on-device models, limited to partners, or your own) to execute plans physically. It's a closed preview, so availability and behavior may change, and pricing specifics aren't published. Its benchmark and safety claims are vendor-reported and early; independent robotics evaluation is immature. And embodied deployment carries real-world safety obligations that go far beyond software testing. Treat it as a powerful preview to prototype with, not a finished production controller.

FAQ

What is Gemini Robotics ER 2?

Google DeepMind's embodied-reasoning vision-language model — a robot's high-level perception-and-planning brain — released in public preview on July 30, 2026, with a real-time streaming variant.

Does ER 2 control robot motors directly?

No. ER 2 is the reasoning/planning layer; direct motor control is handled by separate vision-language-action (VLA) and on-device models, currently limited to early partners.

What can ER 2 do?

Video understanding and moment localization, spatial reasoning, multi-step tool orchestration, progress classification, multi-robot coordination, self-correction, and real-time interaction (streaming variant).

How is ER 2 different from ER 1.6?

Google reports better multi-robot coordination and progress tracking, stronger tool orchestration, and improved safety — including gains on the ASIMOV2 safety benchmark — positioning ER 2 as its safest robotics model yet.

How do I access Gemini Robotics ER 2?

Through the Gemini API (code>gemini-robotics-er-2-preview/code>, code>gemini-robotics-er-2-streaming-preview/code>) and Google AI Studio, as a closed public preview. Pricing follows standard Gemini API billing.

Is ER 2 safe for real robots?

Google positions it as its safest robotics model on safety-constraint adherence and human-proximity, but embodied deployment carries real-world safety obligations; treat safety claims as early and validate rigorously.

Bottom line

Gemini Robotics ER 2 is a major step for embodied AI: a safety-focused reasoning brain that lets robots understand video, reason about space and time, plan multi-minute tasks, coordinate with other robots, and interact in real time via its streaming variant. It's a planning layer, not a motor controller, and it's an early closed preview with vendor-reported benchmarks — but the coordination, progress-tracking, and safety gains over ER 1.6 are meaningful. Prototype with it through the Gemini API for embodiment, and keep the general language/multimodal parts of your stack vendor-neutral through an endpoint like OrcaRouter.