Google DeepMind has officially launched Gemini Robotics ER 2, its most capable embodied reasoning model for robotics to date. Available as of July 30, 2026, the model is accessible through the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform.

What Gemini Robotics ER 2 Actually Does

The model is designed to act as a high-level cognitive layer — or "brain" — for robots, handling reasoning and planning while delegating motor execution to lower-level Vision-Language-Action (VLA) models. This separation of concerns is deliberate: ER 2 focuses on what to do next while the robot simultaneously carries out its current action, eliminating the disruptive "stop-and-think" pauses common in earlier architectures.

Key capabilities include:

  • Continuous video monitoring — the model watches live video feeds to track task progress, detect failures, and decide when to advance to the next step
  • Multi-step task orchestration — developers can declare low-level control interfaces (VLA models, navigation APIs) as callable tools and stream multimodal video, audio, or text directly into the model
  • Native tool calling — including Google Search and user-defined functions, giving robots access to external information mid-task
  • Multi-robot collaboration — enabling multiple robots to coordinate in shared physical spaces on workflows too complex for a single unit

A Meaningful Upgrade Over ER 1.6

DeepMind positions ER 2 as a significant step up from Gemini Robotics ER 1.6. Internal evaluations show ER 2 consistently outperforming its predecessor across three control modes: real VLA, simulation VLA, and human teleoperation.

The model integrates with the Gemini Live API, using a bidirectional streaming endpoint optimized for latency-sensitive real-world tasks. This is critical — physical environments don't pause for a model to finish reasoning, and prior systems that introduced noticeable hesitation created both safety and usability problems.

To demonstrate real-world applicability, DeepMind built a reference integration with Spot, the quadruped robot from partner Boston Dynamics, using ER 2 to orchestrate Spot's native APIs across multi-step physical tasks.

Why This Matters Beyond the Lab

For builders working on physical AI — whether that's warehouse automation, logistics, or service robotics — ER 2 represents a meaningful shift in what's possible without building a full proprietary stack.

The fact that ER 2 is available via a public API means robotics startups and enterprise teams can plug a capable reasoning layer into their existing hardware and VLA pipelines without training models from scratch. DeepMind is explicitly targeting developers, sharing configuration examples and prompting patterns to reduce time-to-deployment.

The multi-robot collaboration feature is particularly notable for industrial applications, where coordinated fleets of robots handling complex, multi-stage workflows have historically required extensive custom orchestration logic.

The Competitive Context

DeepMind's move comes as the broader field of physical AI — robots that can reason about and act in unstructured real-world environments — is heating up rapidly. OpenAI has made investments in humanoid robotics companies, Figure AI and Physical Intelligence (Pi) are developing end-to-end action models, and Tesla's Optimus program continues to advance. DeepMind's approach of offering a cloud-hosted reasoning layer that integrates with any hardware differentiates it from vertically integrated competitors.

The API-first strategy also lowers the barrier for startups experimenting with physical AI — teams that can't afford to build foundational models can now access state-of-the-art embodied reasoning and focus engineering effort on hardware, application logic, and go-to-market.