VLA Gallery / All models / Industry models

Gemini Robotics-ER

Google DeepMind · 2025-03

기타Industry modelsNot disclosedOther

VLM that outputs spatial understanding results such as pointing, 3D bboxes, and grasp predictions instead of actions.

Paper

Gemini Robotics: Bringing AI into the Physical WorldarXiv 2503.20020

Overview

Category

Industry models

Type

기타

Released

2025-03

Organization

Google DeepMind

Open weights

No

Parameters

Not disclosed

Architecture

Backbone LLM

Gemini 2.0 Flash

Vision encoder

—

Action head

Points, bboxes, trajectories, and code output instead of actions (Other)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Embodied-reasoning VLM based on Gemini 2.0; no direct action output (outputs 2D points · 3D bboxes · grasps · trajectories, controls robots via code generation)

Training

Pretraining accelerators

TPU v4, v5p, v6e (chip count not disclosed; model card shared with Gemini Robotics)

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed

Data type

Image·text·robot sensor·action data (model card description shared with Gemini Robotics)

Data scale

Not disclosed

Main datasets

Not disclosed

Tasks & evaluation

Proposes ERQA with 400 questions (Table 2 has no column for ER's own score); 2D pointing Paco-LVIS 71.3, Pixmo-Point 49.5, Where2Place 45.0; blog: end-to-end control success 2~3× that of Gemini 2.0

Robot

N/A (robot control done via code generation and in-context examples)

Inference

Inference device & speed

Not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not disclosed · No mention of RL

Base control

Fixed · No direct action output (generates points, bboxes, trajectories, code); no base control

Notes

Key features

VLM that outputs spatial understanding results such as pointing, 3D bboxes, and grasp predictions instead of actions. Introduced the ERQA (400 questions) benchmark; the Gemini Robotics backbone is a distilled version of this model.

Notes

A VLM (ER) that does not output actions directly, so type is Other. The model card Hardware row (TPU v4, v5p, v6e) also applies to ER (corrects 'Not disclosed†' in the notes). Training Data row: 'Gemini Robotics-ER and Gemini Robotics were trained on datasets comprised of images, text, and robot sensor and action data.'

Verification

Corrections applied