Google DeepMind has launched Gemini Robotics On-Device, a robotics model optimized for general-purpose dexterity and quick task adaptation. This model enhances the capabilities of the previous Gemini Robotics VLA, utilizing local operation for robust performance even in environments with limited connectivity, ensuring low-latency response for sensitive applications. DeepMind is also introducing a Gemini Robotics SDK, enabling developers to test and adapt the model in various domains with minimal demonstrations.
Gemini Robotics On-Device has demonstrated strong generalization and dexterous manipulation capabilities across multiple tasks, significantly outpacing previous models in local operations. It can perform tasks like unzipping bags, and has been successfully adapted for various robot types, including bi-arm robots.
Emphasizing responsible development, DeepMind adheres to AI principles, ensuring semantic and physical safety while actively engaging with feedback from the community. This innovation aims to enhance accessibility and adaptability within the robotics field, encouraging developer collaboration to overcome latency challenges.
Source link
