Photo by Anna Shvets on Pexels

Google DeepMind is pushing the boundaries of robotic autonomy with a new, compact version of its Gemini Robotics AI model. This on-device iteration allows robots to operate independently, performing complex tasks without relying on cloud connectivity. The vision-language-action (VLA) model mirrors the capabilities of its predecessor but is optimized to run directly on robot hardware.

This advancement allows robots to seamlessly adapt to new environments, interpret commands, and execute tasks requiring intricate motor skills – even tasks they haven’t been specifically trained for. According to Carolina Parada, Head of Robotics at Google DeepMind, while the original Gemini Robotics model used a hybrid cloud/on-device approach, this new iteration delivers near-flagship performance offline.

The on-device model displays remarkable adaptability, learning from limited demonstrations and immediately executing a wide range of tasks. Initially trained on Google’s ALOHA robot, the model can be generalized to control other robot types, including the Apptronik Apollo humanoid and the Franka FR3 bi-arm robot.

To foster innovation, Google is releasing a software development kit (SDK) for the on-device model. This marks the first time Google DeepMind has made an SDK available for one of its VLAs, enabling developers to test and refine the technology. Currently, the on-device Gemini Robotics model and its SDK are being rolled out to a select group of trusted testers as Google continues to prioritize safety and minimize potential risks before a broader release.