Gemini Robotics ER 2

Google DeepMind, US · July 2026

Reasoning VLAClosed weightsAlso written GR-ER 2, gemini-robotics-er-2-preview, gemini-robotics-er-2-streaming-preview
Parameters
not published
No parameter count published.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
closed, available through the Gemini API

What it is

Gemini Robotics ER 2 is the current embodied reasoning model in the Gemini Robotics family, published on 30 July 2026 and based on Gemini 3.5 Flash. It takes interleaved text, image, video and audio with a 128k context and up to 64k output tokens, and adds video moment finding, video progress classification, multi-robot orchestration and multi-step tool use over ER 1.6. A streaming variant exists for low-latency agents over the Live API. It is the only part of the family that is broadly available to developers, but the model card publishes no numeric benchmark table and Google prohibits use in safety-critical settings such as healthcare and transportation.

Architecture

Backbone
Based on Gemini 3.5 Flash, as stated in the model card. Inputs are text, image, video and audio with interleaving.
Action head
None. It produces text, spatial and temporal reasoning, tool orchestration and success detection for physical agents.
Parameters
No parameter count published. The model card refers to the Gemini 3.5 Flash model card for architecture details.
Pretraining data
Gemini 3.5 training datasets plus additional datasets representing embodied reasoning tasks, per the model card. No corpus size published.
Embodiments
Embodiment agnostic. Positioned for multi-robot orchestration rather than for a specific platform.

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Offered as the reasoning and orchestration layer for physical agents, including multi-robot coordination and multi-step tool use. The model card points to figures in the release blog for evaluation results rather than publishing a numeric table, so no exact scores are quoted here. A safety benchmark, ASIMOV-Agentic, was released on Hugging Face alongside it.

Fine tuning it yourself

No fine-tuning. It is consumed as an API model, gemini-robotics-er-2-preview or gemini-robotics-er-2-streaming-preview, through Google AI Studio, the Gemini API, or the Gemini Enterprise Agent Platform in private preview.

Where it helps, where it does not

Strengths

  • The only Gemini Robotics model line that is generally available through the Gemini API and AI Studio
  • 128k token input context and up to 64k token output
  • Accepts interleaved text, image, video and audio
  • A dedicated streaming variant exists for low-latency agents over the Live API
  • Adds video moment finding, video progress classification and multi-robot orchestration over ER 1.6

Limits

  • No numeric benchmark table is published. The model card defers to figures in the release blog
  • Reasoning only, it produces no motor commands and needs a VLA to act
  • Closed weights, API only, with the enterprise path still in private preview
  • Explicitly not permitted for safety-critical use including healthcare and transportation
  • Preview status, and the predecessor ER 1.6 is being deprecated within months, which signals a fast deprecation cadence for this line

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.