Google DeepMind introduces on-device Gemini AI model for robots - The Robot Report

Google DeepMind has introduced Gemini Robotics On-Device, a new AI model designed for two-armed robots that enables general-purpose dexterity and rapid task adaptation while operating locally with minimal computational resources. This vision language action (VLA) model brings the multimodal reasoning and real-world understanding capabilities of Gemini 2.0 into physical robotics, allowing robots to perform complex tasks such as unzipping bags or folding clothes. The on-device nature of the model ensures low-latency inference and robustness in environments with limited or no network connectivity, making it suitable for latency-sensitive applications. Alongside the model, DeepMind released a software development kit (SDK) to help developers test and fine-tune the model for specific tasks and environments, requiring as few as 50 to 100 demonstrations. DeepMind has demonstrated the adaptability of the Gemini Robotics On-Device model beyond its initial training on ALOHA robots by successfully applying it to other platforms, including the bi-arm Franka FR3 robot and the