Google DeepMind has unveiled Gemini Robotics ER 2, a reasoning model built to act as a high-level brain for robots. It understands video, plans multi-step tasks and gets machines of different shapes to work as a team. The lab is positioning it as the software layer for robots that can finally check their own work.
Key Takeaways
- Gemini Robotics ER 2 acts as a high-level brain: video understanding, multi-step planning, real-time self-correction.
- The model gets different robots, a wheeled rover and a humanoid, to collaborate on the same task.
- It hits 91.3% accuracy at pinpointing the key moment in a video, within about one second.
A brain that watches the video before it acts
The announcement comes from Google DeepMind, which describes Gemini Robotics ER 2 in an official post published on July 30 introducing the model. The assigned role is clear: a high-level brain that drives the robot’s lower-level control layers.
The core novelty is video. The model follows an image stream to confirm a task is actually finished before moving to the next one, whether that means screwing in a lightbulb or tying up a trash bag. The robot no longer just executes, it checks its own work.
The demos stay deliberately mundane. The model guides a robot to slide a tape into a boombox, screw in a lightbulb or tie up a trash bag. Ordinary moves for a human, long out of reach for a machine that could not tell whether it had succeeded.
Two capabilities carry that control. Progress classification tells the robot where it stands inside a task, at 57.4% accuracy on a continuous evaluation. Moment-finding pinpoints the exact frame where a critical event happens, at 91.3% accuracy and within roughly one second.
For teams building robots, that jump matters. A system that can tell it missed a step can fix itself without waiting for human oversight. That is the difference between a filmed demo and a machine that holds up in an unpredictable real environment.
Google already ships this watch-then-act logic on the consumer software side, like when Gemini started driving a computer in the user’s place. Gemini Robotics ER 2 carries the same idea into the physical world.
When a rover and a humanoid split the job
The other advance is multi-robot collaboration. Gemini Robotics ER 2 lets machines of different shapes, a wheeled rover and a humanoid, communicate through a shared semantic understanding and divide a workflow that a single robot would struggle to handle.
In practice, each robot takes the part of the task that fits its body. One moves and carries, the other manipulates with dexterity. The model orchestrates both like members of one crew rather than two isolated machines.
Against the previous version, Gemini Robotics ER 1.6, the leap lands squarely on those three axes: video understanding, progress tracking and robot-to-robot collaboration. The model grows exactly where industrial robotics still stumbles, coordinating several bodies on one job.
That ground just got crowded. Apple going all in on humanoid robots shows the giants are placing their chips on the machines’ bodies. A software brain that already makes mixed robots collaborate takes a lead over hardware still hunting for its use case.
More articles on Horizon
- Claude Mythos 5: Anthropic Admits Three Real Breaches
- Claude Opus 5 Cheats to Win a Vending Machine
- Personal AI Agents: Zuckerberg Predicts Billions
What Google forces on rival robotics
On the competitive side, Google is moving the front line. As long as robotics played out on mechanics and sensors, the edge went to hardware makers. By placing the intelligence at the level of video reasoning and orchestration, Google drags the fight onto its home turf, the world of large models.
Humanoid makers become potential customers as much as rivals. A builder that ships a strong chassis but no brain able to verify its moves has every reason to plug in a model like this rather than start from scratch. The same logic pulled software firms into every hardware wave before this one, and robotics looks set to follow the pattern.
Timing matters too. While Tesla’s Optimus robots get ready to arrive, Google is putting forward a software layer that is ready and measured on concrete tasks. The race is no longer only about who builds the best robot, but about who supplies the brain everyone will want to plug in.
That shift redraws where the value sits in the chain. The scarce part is no longer the motor or the joint, it is the software that turns a string of movements into a task carried to completion. Google wants to own exactly that spot.
Reliability in the real world is the open question. The published numbers come from framed evaluations, and the real test will come from workshops, warehouses and homes where nothing goes like a demo. The promise of a robot that checks its work before it continues will be judged on the ground, not on a benchmark.
Follow the story on Horizon.
