Google DeepMind Unleashes AI Model for Direct Robot Integration

Google DeepMind is introducing an on-device version of its Gemini Robotics AI model, designed to function without internet connectivity. This vision-language-action model can efficiently operate on robots, enabling them to tackle various tasks, even those not specifically trained for, thanks to its ability to generalize and adapt to new situations. According to Carolina Parada, head of robotics at DeepMind, this model retains many capabilities similar to the flagship version but is optimized for offline use, making it suitable for environments with limited connectivity. Initially trained on the ALOHA robot, it has also been successfully adapted for other robot types like the humanoid Apollo and bi-arm Franka FR3. While the hybrid model remains more powerful, Parada expresses surprise at the on-device model’s performance. Accompanying the launch is a software development kit (SDK) for developers to customize the model, initially available to select testers as Google prioritizes safety.

Source link