Gemini Robotics On-Device
2025-06-24 · Publisher : Google · Domain : Artificial Intelligence
Google DeepMind introduces Gemini Robotics On-Device, its vision-language-action model optimized to run on the robot itself, adaptable to new tasks with 50 to 100 demonstrations.
What changes
- Model executed locally on the robot, with minimal compute requirements, designed for two-armed robots.
- First Google VLA model offered for fine-tuning.
- Adaptation to new tasks with only 50 to 100 demonstrations.
- Highly dexterous tasks performed without a connection: unzipping a bag, folding clothes.
- Gemini Robotics SDK with a MuJoCo simulator, accessible through a trusted tester program.
### Specifications - Type: vision-language-action model for bi-arm robots, on-board execution. - Availability: trusted testers (SDK on registration). ### Pricing Pricing not disclosed at launch. ### Benchmarks The post presents charts of visual, semantic and behavioral generalization against the flagship model and the best previous on-device model, with no figures in the text. ### Key takeaways A step toward autonomous robots with no dependence on the cloud. Access restricted to a small circle in 2025.
Links
Tags : LLM, Google, DeepMind, Robotique, Embarqué