The robot that knows when to stop

The next robot skill may come from watching the whole task.

Google DeepMind's Gemini Robotics ER 2, released on July 30, is a high-level reasoning model for robots. It watches continuous video, plans several steps, and passes the next move to a lower-level action model or robot API. The same layer can coordinate a real robot, a simulated robot, or a human tele-operator.1

The physical detail matters. DeepMind reports 91.3% accuracy for finding the exact moment when a task changes state, with a 0.96-second mean absolute distance. The release also reports 4× execution speed versus larger model categories. A Boston Dynamics Spot demo uses natural language to fetch a snack, while the model tracks progress and decides when the task is complete.2

The shift is from a robot that follows one action policy to a reasoning layer that can watch, hand off, verify, and recover across a longer workflow.