Google DeepMind has unveiled Gemini Robotics On-Device, a groundbreaking advancement in AI that brings powerful robotics capabilities directly to local machines. This version of Gemini’s Vision-Language-Action (VLA) model is optimized to run independently on physical robots, eliminating the need for constant cloud connectivity.
AI Dexterity at the Edge
Gemini Robotics On-Device introduces a leap in dexterous manipulation and real-time adaptability. By functioning directly on the robot, it enables ultra-low latency control and ensures operations continue even when the internet connection drops—ideal for real-world industrial, remote, or sensitive environments.
Key Features and Capabilities
- Multimodal Understanding: Combines visual, linguistic, and behavioral data to complete complex tasks like folding clothes or zipping bags.
- Low-Latency Inference: Executes commands quickly without relying on external servers.
- Adaptability: Fine-tunes with just 50–100 demonstrations to learn new tasks and workflows.
Gemini Robotics On-Device shows remarkable generalization across various scenarios and tasks. It not only understands and responds to natural language instructions but also executes multi-step, high-precision manipulations on bi-arm robots.
Powerful SDK for Developers
To accelerate adoption, DeepMind has released the Gemini Robotics SDK. This toolkit enables developers to test the model in simulation using MuJoCo physics environments and adapt it to new domains efficiently.
The model can even be fine-tuned for different robotic platforms like the bi-arm Franka FR3 and Apptronik’s Apollo humanoid, showcasing its flexibility across embodiments.
Outperforming the Competition
Compared to previous on-device models, Gemini Robotics On-Device shines in both instruction-following and out-of-distribution task performance. It not only achieves parity with its cloud-based counterpart but also outperforms other local models in real-world scenarios.
For a broader perspective on the evolution of Gemini models, see our coverage of Gemini 2.5’s expansion with Flash-Lite and Pro upgrades.
Built With Safety in Mind
Google DeepMind is following a responsible approach to deployment by aligning development with their AI Principles. They recommend rigorous testing using semantic safety benchmarks like Asimov and red-teaming exercises to ensure robustness and compliance in real-world use cases.
Accelerating Robotics Innovation
This launch signifies a key milestone in democratizing high-performance AI for robotics. By enabling local deployment, Gemini Robotics On-Device addresses latency, connectivity, and performance challenges—paving the way for smarter, more autonomous machines.
To get started, developers can sign up for the trusted tester program and explore the full capabilities of this adaptable model.
For an in-depth technical breakdown, check out the Gemini Robotics tech report.
