Google DeepMind · launched July 30, 2026 Public preview · Gemini API and AI Studio

Quick answer

Gemini Robotics ER 2 plans. Your control stack moves.

Use ER 2 as a high-level embodied-reasoning layer: stream or upload physical-world context, plan work, call bounded tools, track progress, and hand execution to a VLA model or robot API. Keep motor control and safety-rated behavior outside the model.

New sibling layer · August 2026

ER 2 is not a hardware-driver standard

Gemini Robotics ER 2 interprets multimodal context, plans, calls declared tools, and tracks progress. Anthropic's Model Hardware Standard is a separate, model-agnostic research preview for making programmable devices discoverable through standardized drivers and read/write primitives, with access through MCP, CLI, or APIs.

A team could evaluate an MHS-backed device capability as one bounded tool beneath ER 2, but neither provider source establishes native compatibility. Keep device identity, command authorization, deterministic limits, controller behavior, independent interlocks, expert oversight, and real-hardware validation outside the model.

Use the MHS versus MCP architecture guide for the layer map and the MHS readiness check before designing a combined proof of concept.

Verified model record

Role

High-level embodied reasoning: understand a physical scene, plan a multi-step task, call tools, and hand motor execution to a lower-level VLA model or robot API.

Standard endpoint

`gemini-robotics-er-2-preview` for spatial reasoning, video understanding, structured output, code execution, grounding, and tool orchestration.

Streaming endpoint

`gemini-robotics-er-2-streaming-preview` for low-latency Live API sessions with continuous multimodal input and blocking function calls.

Input and output

Both preview endpoints accept text, images, video, and audio and return text. They do not directly return motor commands unless your application exposes a structured robot tool.

Published access

Public preview in Google AI Studio and the Gemini API; private preview on Gemini Enterprise Agent Platform.

Flowith access

Not established by Google's launch. Check the current Flowith model selector independently.

Layer decision

Separate reasoning, streaming, and motor execution

Analysis and planned orchestration

Standard preview

Use when structured output, code execution, file or URL context, batch work, or richer non-streaming analysis matters.

Reactive robot-agent loop

Streaming preview

Use when a persistent Live API session, continuous audio or camera input, and low-latency function calling matter.

Motor execution

VLA or robot API

Keep navigation, grasping, movement, and actuator control behind explicit tools, independent safety controls, and validated hardware adapters.

Physical-system boundary

Model safety is not machine safety

Google reports improved instruction-following and human-proximity evaluations. Those results do not certify a robot, site, task, actuator, sensor stack, or operating procedure.

  1. 01 Do not use the model for safety-critical healthcare, transportation, or other work where failure could cause death, injury, or property damage.
  2. 02 Keep collision avoidance, limits, interlocks, geofencing, watchdogs, and emergency stop outside the model.
  3. 03 Restrict tools by name, schema, workspace, robot state, velocity, force, duration, and authorization.
  4. 04 Require notice and consent before identifiable people enter camera or microphone range, and minimize personal-data capture.
  5. 05 Test in simulation and an isolated physical cell before supervised pilots; define safe stop, timeout, rollback, and human takeover.
  6. 06 Treat preview endpoints, model behavior, pricing, availability, and terms as changeable dependencies.

Continue by decision intent

API access guide Choose standard or streaming preview, secure credentials, verify pricing, and plan migration. Robot-agent architecture Connect sensors, reasoning, tools, control adapters, responses, and independent safety gates. Safety and progress Evaluate progress classification, moment finding, intervention, safe stop, and personal-data boundaries. Robostral Navigate Compare a narrower single-camera navigation model with different access and control boundaries. Model Hardware Standard Separate ER 2 reasoning and API tools from a model-agnostic programmable-device driver layer.

Official sources

  1. Google launch — role, tool orchestration, progress tracking, collaboration, access, and provider evaluations.
  2. Gemini Robotics ER API overview — endpoint IDs, input and output types, capabilities, limits, terms, and migration.
  3. Streaming guide — Live API protocol, input formats, blocking tools, and agent loop.
  4. Model card — limitations, safety evaluations, and prohibited safety-critical use.
  5. Anthropic MHS preview — separate model-agnostic device-driver layer, current preview status, access mechanisms, and safety boundary.

Gemini Robotics ER 2 questions, answered

Gemini Robotics ER 2 is Google's embodied-reasoning model for understanding physical scenes, planning multi-step work, tracking progress, calling tools, and coordinating robots. It is a high-level reasoning layer, not a universal motor controller.
Google documents gemini-robotics-er-2-preview for standard interactions and gemini-robotics-er-2-streaming-preview for low-latency Live API sessions. Both are preview endpoints and can change.
Not by itself. Your application declares robot APIs, navigation systems, or lower-level VLA models as tools, executes authorized calls through a hardware adapter, and returns the result to the model.
No. Google's model card says not to use the Robotics Models for safety-critical applications or work such as healthcare, transportation, or other settings where malfunction could foreseeably cause death, injury, or property damage.
Google lists public-preview access through Google AI Studio and the Gemini API, plus private preview on Gemini Enterprise Agent Platform. Verify current account, region, price, quota, terms, and endpoint status.
This page does not claim a Flowith mapping. Google's developer availability and the current Flowith model selector are separate sources, so check the workspace directly.
No. ER 2 is Google's embodied-reasoning model and API layer. Anthropic describes MHS as a separate, model-agnostic research-preview specification for programmable-device drivers accessed through MCP, CLI, or APIs. No documented ER 2 and MHS compatibility is assumed here.