Google DeepMind's Gemini Robotics ER 2 gives bots a real-time brain

If you have ever watched a robot try to clean a messy living room, you know the most painful part is the hesitation. It picks up a cup, freezes for three seconds to think, takes a step, freezes again, and then pours water right over the rim because it missed the exact moment to stop.

Google DeepMind just dropped Gemini Robotics ER 2, targeting that exact disconnect between reasoning and the real-time physical world.

ER 2 serves as the high-level brain in DeepMind's refreshed robotics lineup. While the heavy-lifting Vision-Language-Action models handle real-time motor control and on-device models adapt to new hardware within hours, ER 2 orchestrates the entire mission.

Real-time video tracking and no more awkward pauses

The biggest leap in ER 2 comes down to temporal awareness. Robots usually struggle to know when a task is finished, like whether a trash bag is actually tied or a screw is properly tightened.

DeepMind introduced continuous progress classification, breaking video feeds into five progress buckets from zero to one hundred percent. The model monitors the stream live with 57.4% accuracy, letting the robot self-correct mid-stride without rebooting the whole routine if it drops an item.

It also hits 91.3% accuracy on moment-finding tasks, recognizing the exact frame an action finishes with an average error of under one second. Crucially, it runs at four times the execution speed of larger frontier models.

By routing through the Gemini Live API with bidirectional streaming, ER 2 can plan the next step while the robot is already in motion. DeepMind showed this off with Boston Dynamics' Spot, having the robot dog seamlessly navigate and retrieve a bag of popcorn entirely from natural language instructions.

Humanoids, robot dogs, and team play

DeepMind also demonstrated multi-robot collaboration. Because ER 2 operates as a semantic reasoning engine, machines with completely different bodies can coordinate without hard-coded network protocols.

In one demo, Apptronik's Apollo 2 humanoid robot worked alongside a dual-arm Franka setup. Apollo used its 22-degree-of-freedom SharpaWave hand to manipulate delicate items like ziplock bags, while the Franka arms managed tight packing.

Safety logic is baked straight into the loop. If a human walks into Apollo's work zone, the robot stops immediately and resumes only after the path clears.

Developers can already access ER 2 through Google AI Studio and the Gemini API, while the low-level action and on-device models remain in preview for select partners.

Why it matters

We think DeepMind's decision to decouple reasoning from physical actuation is the right architectural bet. Asking a massive frontier model to output joint angles at 50Hz is an expensive nightmare. Letting ER 2 act as an orchestrator over lighter action models keeps compute costs realistic while giving robots legitimate common sense.

For robotics builders, this turns high-level autonomy into an API call. Instead of hand-crafting brittle state machines to check whether a valve is closed, you stream a camera feed to Gemini and let it return the state.

That said, a 57.4% progress tracking score means the robot will misjudge completion four times out of ten on complex tasks. That is not nearly reliable enough for unsupervised industrial assembly or kitchen cooking. The interface is ready, but the margin of physical error still needs tightening.