Google DeepMind is rolling out its next-generation Gemini Robotics model family. There's Gemini Robotics 2 for controlling movement, ER 2 for understanding environments and planning tasks, and On-Device 2, which runs directly on robot hardware.
Gemini Robotics 2 turns visual input and text commands into actions. The same model checkpoint can drive humanoid robots, dual-arm robots, and various robotic hands — no need to build a separate model pipeline for each machine.
It also gains whole-body control. Humanoid robots can walk, squat, maintain balance, manipulate objects, and handle delicate tasks like tying knots and sealing bags. ER 2, meanwhile, is responsible for breaking down complex tasks, tracking progress, and coordinating multiple robots working together.
Right now, ER 2 is available in Google AI Studio and has entered private preview on the Gemini Enterprise Agent Platform. Gemini Robotics 2 and On-Device 2 remain restricted to early partners.