Gemini Robotics-ER 1.5

Google DeepMind, US · September 2025

Reasoning VLAClosed weightsAlso written GR-ER 1.5, gemini-robotics-er-1.5-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 1.5 is the embodied reasoning model of the 1.5 family, built on the Gemini 2.5 generation, and the first Gemini Robotics model that ordinary developers could call through the API. Its strongest claim is aggregate rather than per benchmark: an ER Score of 59.6 over 15 academic embodied reasoning benchmarks against 51.1 for GPT-5, with an overall average of 62.8 against 60.6. Looked at individually the picture is mixed, since GPT-5 leads on ERQA and the older ER 1.0 leads on Point-Bench. Its reasoning depends heavily on the thinking budget, and disabling thinking costs roughly six aggregate points. It has since been replaced in the API by ER 1.6 and ER 2.

Architecture

Backbone
Built on the Gemini 2.5 generation. A vision-language model for embodied reasoning with a configurable thinking budget, not a motor-control model.
Action head
None. It outputs text, points, plans, progress estimates and tool calls, and acts as the orchestrator for a VLA such as Gemini Robotics 1.5.
Parameters
No parameter count published.
Pretraining data
Gemini 2.5 training data plus additional embodied reasoning datasets. No corpus size published.
Embodiments
Embodiment agnostic. Evaluated as the orchestrator for ALOHA, bi-arm Franka and the Apollo humanoid.

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.

Embodied reasoning

Offline benchmarks that test understanding rather than control. No robot moves in these, so a high score says nothing about grasping.

  • Point-Bench pointing
    71.6%
    accuracyTable 19. GPT-5 scores 43.6. Gemini Robotics-ER 1.0 scores 75.7, so ER 1.5 is slightly below its predecessor on this split.source
  • 15 academic embodied reasoning benchmarks Overall Average
    62.8 points
    aggregate scoreTable 19. GPT-5 60.6, Gemini 2.5 Pro 59.3. Baselines were accessed between 1 and 20 September 2025 with default thinking budgets and no tool use.source
  • 15 academic embodied reasoning benchmarks ER Score aggregate
    59.6 points
    aggregate scoreTable 19. GPT-5 51.1, Gemini 2.5 Pro 51.7, Gemini Robotics-ER 1.0 52.6. This is the split where ER 1.5 leads clearly.source
  • Where2Place free-space pointing
    59%
    accuracyTable 19. Gemini 2.5 Pro 37.0, GPT-5 37.0.source
  • ERQA 400 multiple-choice embodied reasoning questions, thinking enabled
    54.8%
    accuracyTable 19. GPT-5 scores 59.0 and Gemini 2.5 Pro 56.0 on the same split, so ER 1.5 does not lead on ERQA itself. Without thinking it drops to 47.0.source
  • RefSpatial spatial referring
    48.5%
    accuracyTable 19. GPT-5 23.5.source

On real hardware

Used as the high-level orchestrator in an agentic system with Gemini Robotics 1.5 as the low-level controller, evaluated on ALOHA, bi-arm Franka and the Apollo humanoid. The report shows the agent with ER 1.5 as orchestrator beating the same agent driven by off-the-shelf Gemini 2.5 Flash across all tested cases.

Fine tuning it yourself

No fine-tuning. The model was served as gemini-robotics-er-1.5-preview through the Gemini API and Google AI Studio, with behavior controlled by prompting and the thinking_config option, which sets or disables the thinking budget.

Where it helps, where it does not

Strengths

  • Leads the 15-benchmark aggregate (ER Score 59.6 versus 51.1 for GPT-5)
  • Was the first Gemini Robotics model made broadly available to developers through the Gemini API
  • Configurable thinking budget lets you trade latency against reasoning quality
  • Native tool calling, including Google Search, and semantic safety reasoning

Limits

  • Does not lead every benchmark. GPT-5 beats it on ERQA (59.0 versus 54.8) and ER 1.0 beats it on Point-Bench (75.7 versus 71.6)
  • Thinking is load bearing. Turning it off drops the aggregate from 62.8 to 56.7
  • Produces no motor commands, so it needs a separate VLA to act
  • Closed weights
  • No longer listed in the current Gemini API model lineup, which now offers ER 2 and the deprecating ER 1.6

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.