We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
DeepMind Blog·Jun 24·5 min read
Gemini Robotics On-Device brings AI to local robotic devices
TL;DR
Robotic devices often struggle with real-time task execution due to reliance on external data networks. Gemini Robotics On-Device is a vision language action (VLA) model designed to operate locally on robots, enhancing their dexterity and adaptability.
✦ Why It Matters
Engineers can leverage Gemini Robotics On-Device to create more autonomous and adaptable robotic systems in various environments.
Key Takeaways
- Utilize the Gemini Robotics SDK for rapid model evaluation.
- Adapt the model to new tasks with minimal demonstrations.
- Explore its capabilities in environments with limited connectivity.
- Participate in the trusted tester program for early access.
Full Summary
Gemini Robotics On-Device is a new robotics model from Google DeepMind, designed to run locally on robotic devices, enhancing their dexterity and adaptability. This model builds on the capabilities of Gemini Robotics, a vision-language-action (VLA) model, and is optimized for low-latency performance, making it suitable for environments with limited connectivity.
Developers can utilize the Gemini Robotics SDK to test and adapt the model with minimal demonstrations, requiring only 50 to 100 examples for effective fine-tuning. In evaluations, the model demonstrated superior performance in following complex instructions and executing dexterous tasks, such as folding clothes and zipping bags.
It can be adapted to various robotic embodiments, showcasing its versatility across different platforms. This innovation aims to accelerate robotics development by providing a robust tool for developers to create advanced robotic applications.