Multi-Robot Coordination
Robots get a sharedbrain for teamwork.
On July 30, 2026, Google DeepMind announced "Gemini Robotics 2," a family of industrial robotics AI models. At its core is "Gemini Robotics ER 2," the model responsible for coordinating multiple robots. Here's how it lets robots from different manufacturers split up work without a fixed script.
From individually programmed machines to shared understanding
Industrial robots have long meant one machine, one job, one program.
Industrial robots have historically been single-purpose machines, individually programmed for a specific job — one robot for welding, another for material handling. Even when several robots shared a floor, each simply ran its own predetermined sequence; they could not dynamically hand off work to each other based on the situation.
Google DeepMind's "Gemini Robotics 2," announced on July 30, 2026, is a model family built to break that pattern. It consists of three models, and its "high-level brain" is Gemini Robotics ER 2. Beyond real-time spatial reasoning, multi-step task planning, and continuous progress monitoring, it can let different types of robots collaborate through a shared semantic understanding, dynamically allocating subtasks on the fly rather than following a fixed script.
This didn't come out of nowhere. In mid-May 2026, DeepMind partnered with industrial robotics giant Fanuc, announcing plans to integrate Gemini Enterprise and Google's Intrinsic platform across Fanuc's 1.1 million installed robots. Fanuc is also part of DeepMind's "Gemini Robotics Trusted Tester Program," so this ER 2 launch reads less like a one-off demo and more like the continuation of a months-long build-up.
Robots divide labor without a script
In the demo, a humanoid robot and a robotic arm tidied the same room in parallel.
At launch, DeepMind showed a humanoid robot, "Apptronik Apollo 2," and a robotic arm, "Franka F3 Duo," tidying the same room in parallel. ER 2 assigns crouching, pick-up-style tasks to the humanoid and tabletop work to the arm — allocating subtasks on the fly based on which robot is better suited. DeepMind says this already runs on real hardware today from Apptronik as well as Franka and Agile Robotics.
Underpinning this is the Gemini Live API's bidirectional streaming. A robot can keep reasoning about its next move without stopping its current action. There's no "act, then pause to think, then act again" delay — a rhythm that matters when multiple robots are working the same task simultaneously.
Built for developers and industrial floors
This is not a consumer, single-robot story.
Robotics developer teams
Available today through the Gemini API and Google AI Studio, so smaller teams can start experimenting with multi-robot coordination now, without waiting for a full enterprise rollout.
Operators running mixed-vendor fleets
Factories and warehouses mixing Apptronik-, Franka-, and Agile Robotics-class hardware no longer need bespoke coordination code written for every pairing.
Not for consumers, yet
This launch is framed entirely around enterprise and industrial use — don't expect it to show up in a home robot vacuum any time soon.
Watch enterprise rollout and the next hardware partners
Gemini Robotics ER 2 is publicly available to developers via the Gemini API and Google AI Studio, while larger-scale deployment through the Gemini Enterprise Agent Platform remains in private preview.
Evaluate via API/AI Studio first
Teams weighing multi-robot coordination shouldn't wait for general availability on the Enterprise platform — check public access through the Gemini API and AI Studio now.
Track the partner list
Real-hardware support currently spans Apptronik, Franka, and Agile Robotics. Whether that list expands to other manufacturers is the real test of whether this becomes a general coordination layer.
"Its safest robotics model to date" —
a claim made by DeepMind itself.
"Safe" is still a self-reported claim
DeepMind positions ER 2 as its safest robotics model to date on safety-constraint-following and human-proximity benchmarks, saying it can detect nearby humans and trigger a safe stop. Alongside the launch, DeepMind introduced ASIMOV-Agentic, a benchmark measuring whether the model refuses unsafe tool calls and proactively asks for human intervention when it's uncertain a task is safe. But both of these are defined and self-graded by the same lab that built the model. Independent, real-world validation hasn't happened yet, so it's too early to take the "safest" label at face value.