Quick answer
Gemini Robotics ER 2 is publicly available to developers in preview through Google AI Studio and the Gemini API. Choose the endpoint by interaction pattern:
| Need | Model ID | API surface |
|---|---|---|
| Spatial reasoning, video analysis, progress review, structured output, code execution, or planned tool use | gemini-robotics-er-2-preview | Gemini Interactions API |
| Low-latency, continuous robot-agent sessions | gemini-robotics-er-2-streaming-preview | Gemini Live API |
Do not treat the two endpoints as interchangeable. The standard preview supports capabilities that the streaming preview does not, including structured output and code execution. The streaming preview uniquely supports the Live API.
This page was checked against Google’s live documentation on July 31, 2026. Both endpoints are previews; model codes, capabilities, regions, quotas, prices, and terms can change.
Access surfaces
Google names three access paths:
- Google AI Studio: public preview for trying the model and managing an API key.
- Gemini API: public preview for application integration.
- Gemini Enterprise Agent Platform: private preview for an enterprise agent surface.
These do not establish availability in a consumer Gemini app, Google Workspace plan, Google Cloud service not named by the launch, or Flowith.
Start with the Gemini Robotics ER 2 model guide if you still need to decide what the model does. Continue to the robot-agent architecture guide when you are ready to design tools and control boundaries.
Standard preview endpoint
Use gemini-robotics-er-2-preview for interactions where the application can send text, images, video, or audio and receive text after the model reasons over the input.
Google currently documents:
- 131,072 input tokens and 65,536 output tokens;
- text, image, video, and audio inputs with text output;
- function calling and structured output;
- code execution, computer use, file search, URL context, and search or Maps grounding;
- thinking controls and caching;
- Batch API support;
- no Live API support on this endpoint.
The standard endpoint is a better starting point for uploaded video review, spatial analysis, progress classification, moment finding, structured coordinates, or a planned sequence where sub-second reaction is not required.
Streaming preview endpoint
Use gemini-robotics-er-2-streaming-preview for a persistent Live API session that must react to text, audio, and camera input with low-latency tool calls.
Google’s current streaming guide specifies:
- a stateful WebSocket session;
- raw 16-bit PCM audio at 16 kHz, little-endian;
- JPEG camera frames at no more than one frame per second;
- text output;
- function calling with physical actions declared as
BLOCKING; - no structured output, code execution, file search, URL context, caching, or Batch API support on the streaming endpoint.
The endpoint returns text. Speech requires a separate text-to-speech path, and physical execution requires your own authorized tool adapter.
Secure the API path
Google says the robotics API rejects unrestricted API keys with a 403 Forbidden response. Apply restrictions in AI Studio before testing.
For a production-shaped integration:
- store the key in a server-side secret manager;
- restrict it to the intended project and service;
- never expose it in a browser, mobile bundle, log, prompt, or repository;
- isolate development, staging, and production credentials;
- cap spend and quota where supported;
- log model ID, endpoint, request ID, tool calls, latency, and outcome without storing unnecessary personal data;
- rotate on a schedule and on suspected exposure.
Key restriction is only one control. Tool authorization and physical-system safety belong to the robot application.
A minimal standard interaction
Google’s current examples use the Interactions API:
- create a Gemini client for an authorized project;
- upload or reference the media input;
- send an interaction using
gemini-robotics-er-2-preview; - request a bounded output such as labeled coordinates or a progress classification;
- parse and validate the returned text or structured result;
- pass approved output to a robot API or lower-level VLA only through a constrained adapter.
Do not copy an output coordinate directly into an actuator command. Convert coordinate systems, validate scene state, enforce reach and collision limits, and stop safely on uncertainty.
A minimal streaming session
The Live API flow is different:
- declare an allowlisted set of robot capabilities as function tools;
- mark physical actions as
BLOCKING; - connect with
gemini-robotics-er-2-streaming-preview; - stream authorized text, audio, or JPEG frames;
- receive text and tool calls;
- validate a tool call against current robot state and safety controls;
- execute through the robot SDK;
- send a tool response back to the session;
- close or stop on timeout, operator action, state mismatch, or safety event.
Only blocking function calls are currently supported for the robotics Live API. That serializes the model’s tool-selection loop; it does not make the underlying hardware safe.
Migration and lifecycle
Google’s July 30 documentation says ER 1.6 users can replace gemini-robotics-er-1.6-preview with either ER 2 endpoint and that ER 1.6 will shut down at the end of August.
Treat that as a migration trigger, not proof of drop-in compatibility. Re-run:
- spatial and progress evaluations;
- prompt and output-schema tests;
- tool-call authorization;
- latency and timeout measurements;
- safety, privacy, and human-intervention scenarios;
- price and quota checks;
- rollback and endpoint-health procedures.
Store model IDs in configuration. Preview aliases can change behavior even when application code does not.
Pricing and region checks
Google routes the robotics overview to the live Gemini API pricing page instead of publishing a durable robotics price in the model description. Before budgeting, capture:
- model ID and preview status;
- input modality, media duration, and token counts;
- thinking level, grounding, code execution, and tool use;
- cached or batch behavior where supported;
- failed request treatment;
- project region, quota, and account terms.
Do not describe the model as free because AI Studio offers a trial surface or because another Gemini model has a free tier.
Privacy and physical-world data
Google’s robotics terms warn that video and audio may collect identifiable voice, imagery, or likeness data. The documentation requires sufficient notice and consent before identifiable people interact with or enter the area around a Robotics Model, along with reasonable data-minimization efforts such as face blurring and avoiding areas with identifiable people where practical.
Map those requirements into camera placement, microphone range, signage, consent, access, retention, deletion, incident response, and operator training before a live test.
Frequently asked questions
How do I access Gemini Robotics ER 2?
Google lists Gemini Robotics ER 2 in public preview through Google AI Studio and the Gemini API. Create or select an authorized project, restrict the API key as required, choose the correct preview endpoint, and verify current region, quota, price, and terms.
What is the standard Gemini Robotics ER 2 API model ID?
The standard model ID is gemini-robotics-er-2-preview. Google documents it for non-streaming interactions, including spatial reasoning, video understanding, structured output, code execution, grounding, and function calling.
What is the streaming Gemini Robotics ER 2 model ID?
The Live API model ID is gemini-robotics-er-2-streaming-preview. It supports a persistent bidirectional session for text, audio, and camera input with text output and blocking function calls.
Can I reuse an unrestricted Gemini API key?
No. Google’s robotics API documentation says unrestricted API keys receive a 403 response. Restrict the key in AI Studio, keep it out of source control and client bundles, and limit project, service, origin, and rotation scope.
Is Gemini Robotics ER 2 generally available?
No. Google labels both documented API endpoints as preview. Build for model, capability, quota, price, region, and lifecycle changes, and check the live documentation before production work.
Does Gemini API access include Flowith access?
No automatic mapping is established. Google AI Studio, the Gemini API, Gemini Enterprise Agent Platform, and Flowith are separate product surfaces with separate availability and billing checks.
Bottom line
Use the standard preview for deeper non-streaming reasoning and the streaming preview for a reactive Live API loop. Keep endpoint identity, capability checks, restricted credentials, privacy, pricing, lifecycle, motor control, and Flowith availability as separate decisions.