Gemini Robotics On-Device brings AI to local robotic devices

10 months ago

Gemini Robotics On-Device is introduced as an efficient, on-device robotics model with general-purpose dexterity and fast task adaptation.
The model operates independently of a data network, making it suitable for latency-sensitive applications and environments with intermittent or zero connectivity.
A Gemini Robotics SDK is available for developers to evaluate, test, and adapt the model to new domains with minimal demonstrations.
Gemini Robotics On-Device is optimized for bi-arm robots, offering rapid experimentation, adaptability, and low-latency inference.
The model excels in visual, semantic, and behavioral generalization, following natural language instructions and performing dexterous tasks like unzipping bags or folding clothes.
It outperforms other on-device alternatives in challenging tasks and complex multi-step instructions.
Developers can fine-tune the model for better performance, requiring as few as 50 to 100 demonstrations.
The model has been adapted to different robot embodiments, including the bi-arm Franka FR3 robot and the Apollo humanoid robot.
Responsible development and safety measures are in place, including semantic and physical safety evaluations.
Gemini Robotics On-Device is initially released to a select group of trusted testers for feedback and safety evaluation.
The model and SDK aim to accelerate innovation in robotics by addressing latency and connectivity challenges.

Hasty Briefsbeta