Gemini Robotics ER 2: Google DeepMind Teaches Robots to Reason and Work as a Team

Google DeepMind has unveiled Gemini Robotics ER 2, a next-generation model for controlling robots in the real world. Its key capabilities are video-stream understanding, orchestration of multi-step tasks and coordinated work by several robots at once.

What Gemini Robotics ER 2 can do

Gemini Robotics ER 2 is built around three areas, each of which Google DeepMind calls a fundamental step forward for robotics applications.

Video understanding: the model analyzes visual context from an ordinary real-time video stream. It recognizes objects, tracks motion and makes decisions on the fly. A robot no longer needs a rigid, pre-defined map of its environment or a set of specialized sensors.

Task orchestration: the system plans multi-step scenarios, using several tools in the right order. This moves beyond single commands to goal-oriented planning: the robot sees the whole task and builds the sequence of steps itself.

Robot collaboration: several machines act together. They divide subtasks, align their actions and adapt to changes in real time. According to DeepMind, this is exactly what was missing for the move from controlled demos to real production and logistics scenarios.

• 3 key areas: video understanding, task orchestration, multi-robot collaboration
• Developer: Google DeepMind
• Positioning: real-world tasks, not just lab tests
• Lineup: a successor to the previous-generation Gemini Robotics ER

Why robot collaboration is an unsolved problem

Traditional industrial robots work in isolation: each performs a strictly defined operation following a rigid algorithm. Coordinating several machines in real time demands more: a shared understanding of context, the ability to reassign tasks when conditions change, and awareness of what a partner is doing.

This is where AI models change the equation. Instead of hand-written scripts, they let robots reason about a task and adapt to changes on the go, without being reprogrammed for each new case.

"Gemini Robotics ER 2 helps robots reason, interact and solve real-world tasks.

It is a step forward in video understanding, tool orchestration and collaboration among multiple robots," says the official Google DeepMind blog.

The company says the model targets real-world applications rather than demonstrations in controlled environments. That is a significant claim: most robotics systems today still fall short of the required reliability outside the lab.

Where ER 2 fits in the race for physical AI

Google DeepMind is building Gemini Robotics as a lineup that spans the entire robot control stack, from perception to planning and cooperation. Gemini Robotics ER 2 advances on all three fronts at once, which sets it apart from incremental point updates.

Physical AI is one of the fastest-growing areas of competition among the leading labs. The emphasis is shifting from specialized industrial machines to general-purpose systems that can adapt to new tasks without being rigidly reconfigured for each scenario.

What this means

Gemini Robotics ER 2 signals that Google DeepMind is moving from language and vision AI systems toward physical intelligence: robots that can reason, coordinate and operate in the real world rather than in a lab. If the claimed capabilities hold up in practice, they would significantly broaden the range of robot applications in manufacturing, logistics and related industries.