Robots Atlas>ROBOTS ATLAS
Gemini Robotics On-Device
AI Modelsโ€บGemini

Gemini Robotics On-Device

Family: Gemini
VLA model by Google DeepMind optimized to run locally on robotic devices with low-latency inference.
โณ Previewโณ Limited accessRobotics foundation modelVision-Language-Action model๐Ÿ“ Gemini
Context window
1088 tokens
tokens
Release date
14 April 2026
๐Ÿ”ฌGoogle DeepMindResearch lab
Access:Hostedon-deviceDeployment:Edgeโ˜ Cloud

Overview

Gemini Robotics On-Device is a Vision-Language-Action (VLA) model optimized to run directly on robotic hardware without requiring a persistent cloud connection. It features a very small context window (1088 tokens), reflecting its optimization for real-time, low-latency operation.

It is the first Google DeepMind VLA model made available for fine-tuning by robotics developers via the Gemini Robotics SDK. It takes images, text, and action commands as input and outputs action commands controlling the robot. Unlike Gemini Robotics 1.5, it does not produce text output.

Classification
Robotics foundation modelVision-Language-Action model
Family: Gemini
Applications
Workflow automation
Access & deployment
Hostedon-device
EdgeCloud
Weights: Closed
Key parameters
๐Ÿ“ Context: 1088 tokens
โœ“ Fine-tuning
๐Ÿ“ฅ Input: text, image, action
Robotics
Dexterous manipulationRobot manipulationRobot controlMotion planning

Technical specification

Context window
1088 tokens
tokens
Features:โœ“ Fine-tuning
Modalities
โฌ‡ Input
textimageaction
โฌ† Output
action

Capabilities and applications

Native model capabilities
Image understanding
Category: vision
Multimodal understanding
Category: multimodal
Robotics
Dexterous manipulationRobot manipulationRobot controlMotion planning
Application domains
Workflow automation

Sources and related pages