VLA GalleryAll modelsGallery

VLA Gallery / All models / Industry models

Gemini Robotics On-Device

Google DeepMind · 2025-06

VLAIndustry modelsNot disclosedNot disclosed

Designed for minimal compute so it runs locally on the robot without a data network.

View architecture diagram in the galleryOfficial page ↗

Paper

No paper · Related Gemini Robotics: Bringing AI into the Physical World

Overview

Category
Industry models
Type
VLA
Released
2025-06
Organization
Google DeepMind
Open weights
No
Parameters
Not disclosed

Architecture

Backbone LLM
Not disclosed (lightweight VLA in the Gemini Robotics family)
Vision encoder
—
Action head
Not disclosed (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Not disclosed (introduced only as a lightweight VLA in the Gemini Robotics family)

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Accelerator·time not disclosed (adapts with 50~100 demos; trusted tester SDK, supports MuJoCo simulator evaluation)
Data type
Real robot (ALOHA)
Data scale
Not disclosed
Main datasets
In-house ALOHA data (private)
Tasks & evaluation
7 dexterous tasks (zipping a lunch box, drawing a card, pouring salad dressing, etc.); claims an edge over the previous best on-device model in generalization · instruction following (figures in charts only); Franka FR3 folding · industrial belt assembly
Robot
ALOHA (training); adapted to bi-arm Franka FR3 and Apptronik Apollo humanoid

Inference

Inference device & speed
Runs locally on-device on the robot (no data network needed); hardware·latency figures Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Onboard

Reinforcement learning & base

Reinforcement learning
Not disclosed · Training method details not disclosed
Base control
Fixed · No base·locomotion control (bimanual manipulation model; Apollo adaptation is manipulation only)

Notes

Key features
Designed for minimal compute so it runs locally on the robot without a data network. The first VLA DeepMind offered for fine-tuning, adapting to new tasks with 50~100 demos.
Notes
Blog posted 2025-06-24. No parameter·latency·hardware figures in the blog. 'fewer than 200 examples' on the current model page is 2026 On-Device 2 wording, not this model's value. The GR 1.5 report compares against it as a baseline.
Verification
Confirmed

Other models in this category