AY-ROBOTSAI Teleoperation
Sign Up

Gemini Robotics-ER

Google DeepMind, US · March 2025

Reasoning VLAClosed weightsAlso written Gemini Robotics-ER 1.0, GR-ER
Parameters
not published
No parameter count published for Gemini Robotics-ER or for Gemini 2.0 Flash.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

Gemini Robotics-ER is the embodied reasoning half of the original Gemini Robotics release, a Gemini 2.0 Flash variant post-trained for spatial and temporal understanding. It does perception and planning, not motor control: pointing, grasp prediction, 2D trajectories, 3D boxes and code generation that drives the robot. It set a state of the art on SUN-RGBD at 48.3 AP@15 and roughly doubled Gemini 2.0 Flash on zero-shot ALOHA 2 control, but real-robot zero-shot success is only 25 percent and dexterous tasks such as dress folding fail completely. In-context demonstrations are the intended lever, raising real-robot success to 65 percent. It has been superseded three times and is no longer offered in the API.

Architecture

Backbone
A version of Gemini 2.0 Flash with enhanced embodied reasoning. It is a vision-language model, not a VLA, and outputs text, points, boxes and code rather than motor commands.
Action head
None in the motor-control sense. It controls robots indirectly through zero-shot code generation, 2D pointing, grasp prediction, trajectory prediction and 3D bounding boxes.
Parameters
No parameter count published for Gemini Robotics-ER or for Gemini 2.0 Flash.
Pretraining data
Gemini 2.0 multimodal pretraining plus embodied reasoning post-training. The report does not quantify the embodied post-training corpus.
Embodiments
ALOHA 2 (evaluated in simulation and on the real robot via generated code)

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

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • ALOHA 2 sim task suite in-context learning with 10 demonstrations
    65%
    average success ratesource
  • ALOHA 2 sim task suite zero-shot control via code generation
    53%
    average success rateGemini 2.0 Flash reaches 27 percent on the same suite, so the ER post-training roughly doubles it. Average over 50 trials per task.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • ALOHA 2 real robot in-context learning with 10 demonstrations
    65%
    average success ratesource

Fine tuned tasks

What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.

  • ALOHA 2 real robot zero-shot
    25%
    average success rateFold dress scores 0 percent zero-shot. The report attributes this to the inability to generate precise enough grasps.source

Embodied reasoning

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

  • SUN-RGBD 3D object detection, AP@15
    48.3 AP@15
    average precisionReported as a new state of the art. Gemini 2.0 Flash scores 30.7 and ImVoxelNet 43.7 on an easier 10-category set.source
  • ERQA 400 multiple-choice embodied reasoning questions
    45.3%
    accuracyMeasured by the same team in the Gemini Robotics 1.5 report, Table 19, not in the original Gemini Robotics report. ERQA itself was introduced in the Gemini Robotics report.source

On real hardware

Controls a real ALOHA 2 through generated code rather than learned motor policies. Zero-shot it reaches 25 percent average success on the real robot and 65 percent with 10 in-context demonstrations. It fails outright on dress folding (0 percent zero-shot) because it cannot produce sufficiently precise grasps.

Fine tuning it yourself

No fine-tuning is offered. The supported adaptation path is in-context learning, where roughly 10 demonstrations of observations, actions and language are placed in the prompt, which lifted real ALOHA 2 success from 25 to 65 percent.

Where it helps, where it does not

Strengths

  • State of the art on SUN-RGBD 3D detection at 48.3 AP@15 with open-vocabulary queries instead of a closed category set
  • Roughly doubles Gemini 2.0 Flash on zero-shot ALOHA 2 control (53 versus 27 percent in simulation)
  • In-context demonstrations lift real-robot success from 25 to 65 percent without any weight updates
  • Introduced ERQA, an open benchmark of 400 questions that the field still uses

Limits

  • Not a motor-control model. Dexterous tasks fail because grasps are not precise enough, for example 0 percent on dress folding zero-shot
  • 25 percent zero-shot on real hardware is far below what a trained policy achieves
  • Closed weights, and the 1.0 API preview has since been removed in favor of ER 1.6 and ER 2
  • Superseded by ER 1.5 (September 2025), ER 1.6 (April 2026) and ER 2 (July 2026)

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.

Download the clientDrive a real arm now