MANY MINDEDDeepMind ships a robot planner that watches its own work
Google DeepMind released Gemini Robotics ER 2 on July 30, available immediately through the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform. It is not a controller. It is the layer above one: it talks with people, reasons about a physical scene, breaks a job into steps, then hands motor execution to whatever vision-language-action model or navigation API a developer wires in as a tool. It can also call Google Search or any user-defined function mid-task.
The substance is timing. DeepMind reports 57.4% accuracy at classifying how far along a task is, sorting each video frame into one of five progress bands, and 91.3% accuracy at moment-finding — pinpointing the frame where a critical event happens, such as when to stop pouring — with a mean absolute distance of 0.96 seconds. It runs over the Gemini Live API's streaming endpoint, and DeepMind claims four times the execution speed of larger model categories at a fraction of their compute cost. A new multi-robot mode lets different machines share a semantic picture and hand work between them; demonstrations pair it with Boston Dynamics' Spot and with Apptronik's Apollo 2 alongside a Franka F3 Duo.
Knowing when a task is finished is the unglamorous blocker on robot labour. A machine that cannot tell a tightened bulb from a loose one needs a person watching it. Progress tracking and self-correction are what let one supervisor cover many robots instead of one, and that ratio is the whole cost curve for physical work.
Every benchmark figure here is DeepMind's own, published with the launch, and the demos are demos. Watch for independent evaluation and for behaviour on hardware outside the partner set.
Source: Google DeepMind