Google DeepMind unveils Gemini Robotics 2 for whole-body robot control and multi-robot collaboration

Google DeepMind announced Gemini Robotics 2, a new set of models designed to enable robots to reason through complex, whole-body tasks and collaborate as teams. The system consists of three models: a vision-language-action (VLA) model for converting vision and language inputs into motor control, an embodied reasoning (ER) model for multi-step task planning and human communication, and an on-device VLA model optimized for local deployment and rapid adaptation to new robot embodiments.

The VLA model supports control of full humanoids and bi-arm robots, including fine dexterity for both hands and grippers. The ER model acts as an agent that can plan tasks lasting several minutes and facilitate human-robot interaction. The on-device model enables fast adaptation to new robot bodies with just a few hours of data, supporting deployment in edge environments.

Gemini Robotics 2 was demonstrated across multiple robot embodiments, including the Apptronik Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and the Franka Duo with Robotiq gripper. The models achieved medium to high success rates on whole-body and gripper-based dexterous tasks, though multi-finger dexterous manipulation remains challenging.

The ER model is available on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while the VLA and on-device models are available to early-access partners. Google DeepMind also provided guidance for developers on integrating these models into their hardware via a developer blog.