The AI arm of Google, Google DeepMind unveiled a locally deployable robotics model–Gemini Robotics On-Device. This model is designed to operate efficiently on robotic hardware without reliance on external networks. It demonstrates advanced capabilities in general-purpose dexterity and task adaptation across various use cases.
The model functions entirely on the robot, which reduces latency and maintains operational stability even in environments with unreliable or no internet connectivity. It is intended for bi-arm robotic systems and requires minimal computational power.
Gemini Robotics On-Device extends the functionality of previous Gemini Robotics models by supporting real-time, dexterous task execution, enabling fine-tuning for new tasks, and allowing natural language command interpretation. The system shows consistent performance in handling visually and semantically varied tasks such as manipulating soft objects or executing multi-step instructions.
Evaluations indicate that the model generalizes well across different conditions and outperforms comparable on-device systems in complex scenarios. For developers requiring enhanced capabilities beyond local constraints, an alternative version of the Gemini Robotics model is available.
We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. 
It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments – without needing a constant internet connection. 
pic.twitter.com/1Y21D3cF5t
— Google DeepMind (@GoogleDeepMind) June 24, 2025
Gemini Robotics On-Device Becomes First VLA Model Available For Fine-Tuning, Developed With Emphasis On Safety And Responsible Innovation
Gemini Robotics On-Device represents the first instance of a VLA model from this series made available for fine-tuning. While the model is capable of executing a range of tasks without modification, it can also be adapted to enhance performance in specific applications. Adaptation can be achieved using a relatively small number of demonstrations, typically between 50 and 100, which demonstrates the model’s ability to apply its foundational capabilities to unfamiliar tasks.
The development of all Gemini Robotics models follows a framework aligned with established AI Principles, incorporating a comprehensive safety strategy that addresses both semantic and physical dimensions. Semantic and content-related safety is monitored through the Live API, while low-level safety-critical controllers are integrated to manage the execution of physical actions. A semantic safety benchmark has been introduced for system-wide evaluations, and targeted testing methods, including red-teaming, are advised to identify safety-related weaknesses.
Oversight of the models’ real-world implications is conducted by the Responsible Development & Innovation team, which assesses potential risks and societal effects. These findings are reviewed by the Responsibility & Safety Council, whose recommendations inform the ongoing development of the models, aiming to enhance positive outcomes while reducing potential harm.
The post Google DeepMind Drops Gemini Robotics On-Device, Enabling Localized AI Integration For Robotic Systems appeared first on Metaverse Post.