Google DeepMind on July 30 introduced Gemini Robotics ER 2, an update to its “embodied reasoning” model that acts as a high-level planning layer for robots — reading live video, tracking how far a task has progressed, and coordinating handoffs between multiple machines, according to Google DeepMind.
From snapshots to live video
The previous version, ER 1.6, worked mainly from static images. ER 2 instead treats continuous video as a first-class input, letting a robot watch a scene unfold rather than reason about isolated frames. Two capabilities support this: continuous progress classification, which estimates how much of a task is complete in rough 20% increments, and moment-finding, which locates the exact video frame where a key event happened. Google DeepMind says moment-finding reaches 91.3% accuracy, while progress classification is correct 57.4% of the time — a modest figure, but a clear jump from ER 1.6’s baseline.
Faster planning, more robots working together
The model streams through the Gemini Live API, so it can keep planning its next move while a robot is still executing the last one, rather than pausing between steps. Google DeepMind says this cuts round-trip latency to under a second, about four times faster than before. ER 2 also orchestrates lower-level vision-language-action models — the specialized systems that turn a plan into motor commands — and can call outside tools, such as a search engine, when a task needs more context.
A new multi-robot collaboration feature lets machines from different manufacturers share a single semantic understanding of a task and hand off pieces of it to each other. In a demo, Apptronik’s Apollo 2 humanoid and a Franka F3 Duo arm worked from the same plan, while Boston Dynamics’ Spot fetched objects on spoken command. Cross-brand coordination like this goes further than what Figure AI’s humanoid robots have shown on factory floors so far, where each robot still runs its own separate stack.
Gemini Robotics ER 2 is available now through the Gemini API and Google AI Studio, with a private preview inside the Gemini Enterprise Agent Platform.