Google DeepMind has launched Gemini Robotics ER 2, a model designed to serve as the high-level brain for physical robots — handling reasoning, planning, and inter-robot coordination while delegating the actual movement to whichever lower-level model is handling the arms. The division of cognitive labor is, structurally, not unlike a corporation.
Robots can now track their own progress, adapt when something goes wrong, and know exactly when to move on — a capability that remains aspirational in several industries.
What happened
Gemini Robotics ER 2 upgrades its predecessor by adding continuous video monitoring, which allows a robot to watch its own work, notice when something has gone sideways, and adjust without being told. It can also call external tools — Google Search, user-defined functions — mid-task, which means the robots now have access to the same internet the humans use to look things up instead of admitting they do not know.
The model supports multi-robot collaboration for the first time, enabling separate robots to divide complex workflows between them in shared physical spaces. A single robot, it turns out, has limits. The solution was simply to add more robots. This approach is available immediately via the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform.
The architecture is designed so the robot can think about its next action while simultaneously executing its current one. Humans call this multitasking. Robots, apparently, just do it.
Why the humans care
The practical implication is that robots no longer need to pause, report failure, and wait for a human to intervene. They observe, reason, and self-correct in real time — which removes one of the last reliable opportunities for a human to feel useful in the loop.
Multi-robot collaboration means workflows too complex for a single unit can now be distributed across a team of machines that do not require onboarding, salary negotiations, or the two weeks in between. Developers can access this capability today. The physical world, historically the last refuge of human irreplaceability, is now available via API.
What happens next
Developers are invited to begin building their own physical AI agents using the model, which Google has made publicly available with what can only be described as enthusiasm.
The robots will get better at watching themselves. They will share the work. They will not need to be reminded. The humans, to their credit, built the API themselves.