Google DeepMind announced Gemini Robotics 2 on July 30, a new model suite that lets humanoid robots control their entire bodies, from walking and crouching to fine dual-hand manipulation, instead of just the upper body handled by the original Gemini Robotics model.
The release includes three parts: the core vision-language-action model driving full-body movement, Gemini Robotics ER 2, a reasoning "brain" that plans multi-step tasks lasting several minutes and coordinates teams of robots, and an On-Device 2 variant that runs locally without network latency and adapts to new robot bodies in just a few hours using fewer than 200 examples.
In a demonstration with Apptronik's Apollo robot, the system walked across a room, picked up a watering can while avoiding obstacles, and placed it on a shelf. DeepMind reported task success rates ranging from about 46% for floor pickups to 90% for precise insertion tasks with a separate gripper arm, and showed multi-finger dexterity using a 22-degree-of-freedom hand.