Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind just shipped Gemini Robotics ER 2, and the pitch is refreshingly narrow for once: this isn't about robots moving faster or gripping harder, it's about robots knowing what's happening around them and when a job is actually finished. That sounds mundane until you realize how much robotics has struggled with exactly this — a robot that can screw in a lightbulb but has no clean way to tell if it's actually tight, or a robot that pours coffee but can't say precisely when to stop.

The model works as a high-level reasoning layer, not a motor controller. It sits above whatever vision-language-action model actually moves the arm or the legs, calling those systems as tools while it handles planning, tool use like Google Search, and conversation with a human standing nearby. DeepMind calls this an upgrade from ER 1.6, and the numbers back that framing: 57.4% accuracy on tracking task progress across five completion bands, and 91.3% accuracy on pinpointing the exact video frame where something critical happens, with a mean error of under a second. That precision, DeepMind says, comes at a fraction of the compute cost of larger rival models and runs four times faster — which matters a lot when you're deciding, in real time, whether a robot arm is about to knock something off a table.

The more interesting shift is temporal. Instead of judging a single snapshot, ER 2 watches a continuous video feed and reasons about progress over time, which lets a robot notice it's failing mid-task and self-correct instead of restarting from scratch. DeepMind demoed this with Boston Dynamics' Spot, using ER 2 to chain together navigation and manipulation commands so the robot could fetch a snack on a spoken request, coordinating without the stop-and-think pauses that have made earlier robot demos feel stilted. There's also a multi-robot angle: Apptronik's Apollo 2 humanoid and a Franka F3 Duo arm were shown handing off tasks to each other, on the theory that no single robot body is right for every job — wheels are great indoors, legs handle rubble, and a shared reasoning layer lets them split the work sensibly.

Safety gets real attention here too, and not as an afterthought bullet point. DeepMind says ER 2 got meaningfully better at halting a humanoid robot when a person walks into its space and resuming only once the area clears, and the company is introducing a new benchmark specifically for testing whether a model can act as a safe orchestrator — enforcing constraints, monitoring surroundings, judging what's physically feasible, and asking a human when it's unsure. It's a tacit admission that as these systems get more autonomous and start commanding multiple robots at once, the failure modes get more consequential, not less.

ER 2 is live now through the Gemini API and AI Studio, with an enterprise version in private preview. DeepMind is framing it as infrastructure for other people's robots rather than a robot of its own, which fits the company's broader strategy of selling the brain and letting hardware partners worry about the body.