Google DeepMind debuts Gemini Robotics 2 model series for humanoid robots
SiliconANGLE Maria Deutscher ● Covered by 7 sources
Google DeepMind just launched Gemini Robotics 2, a model family that lets humanoid robots plan and split up multi-step chores together. It can also self-correct mid-task instead of starting over when something goes wrong.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind is betting that the next big unlock in robotics isn't better motors or grippers, but smarter brains sitting on top of them. Today the lab rolled out Gemini Robotics 2, a set of models built specifically for humanoid robots, and the headline feature is coordination: multiple machines can now split a single long chore between them and work on it in parallel.
The architecture follows a pattern that's become standard in humanoid robotics: a high-level reasoning model draws up the plan, and a separate low-level model translates that plan into motor commands. DeepMind's reasoning piece here is called Gemini Robotics ER 2, and it's the part doing most of the heavy lifting. Tell it a task in plain English, even something that takes hundreds of steps and several minutes to finish, and it can break the job down, hand pieces off to other robots, and even reach out to Google Search mid-task if it doesn't understand something in the instructions.
What's more interesting than the multitasking, though, is how ER 2 handles failure. Robots mess up constantly, and until now the fix was usually to restart the whole sequence. ER 2 instead watches continuous camera footage of its own robot in action, figures out the last step that actually succeeded, and resumes from there. Google engineers Steven Hansen and Peng Xu described it as robots that can
My take — AI-written commentary, not fact-checked reporting
Splitting chores across a fleet of robots sounds impressive in a demo, but the real story is the self-correction feature, because that's the boring, unglamorous problem that's actually been holding humanoid robots back in messy real-world settings. Everyone chases the flashy multi-robot coordination headline; the people who actually deploy this stuff will care far more about a robot that doesn't restart from zero every time it drops a mug. Google shipping a safety benchmark alongside the model is the right instinct too, though don't expect much scrutiny until these things start working alongside people in warehouses rather than in curated press footage.
Read more about this at: SiliconANGLE