Robots That Learn Like People?

New Gemini AI Could Change Everything

If you’ve ever watched a video of a robot struggle to pick up a banana or fold a towel, you’ve seen the frustrating gap between what AI can think and what robots can do. Google DeepMind may have just taken a major step toward closing that gap. (see my article “Robots Can Dance”)

Earlier this week, DeepMind unveiled Gemini Robotics—a powerful new AI system that enables robots to complete a wide range of everyday physical tasks with minimal prior training. That’s a big deal. Until now, most robots needed thousands of repetitions to learn simple actions like turning a doorknob or unscrewing a bottle cap. Gemini Robotics does it after watching a short video—or just from reading a description.

The system builds on the company’s Gemini 1.5 family of large language models, blending language understanding, vision, and physical coordination. What makes Gemini Robotics especially exciting is its ability to generalize: it can take what it learns in one situation and apply it to others, without being explicitly reprogrammed. That kind of flexibility is exactly what’s been missing from most robotic systems to date.

In demonstrations, robots powered by Gemini were able to:

• Fold a piece of paper into a paper airplane
• Stack irregular blocks
• Pick up objects from cluttered environments
• Use simple tools like pens, cords, and scissors
• Respond accurately to natural-language instructions like “tie the shoelace” or “push the red button”

And it’s not just a lab curiosity. These kinds of skills—basic motor tasks with everyday materials—are the backbone of household and service robotics, eldercare systems, and even next-gen manufacturing. They’re the reason why we don’t have more useful robots in our homes today.

The takeaway? This isn’t about flashy sci-fi androids. It’s about giving robots the ability to interact with the physical world in fluid, intuitive ways. And that opens the door to real-world applications in logistics, medicine, caregiving, and personal assistance.

Of course, the tech isn’t perfect yet. DeepMind’s team acknowledges that the robots still make errors—and they don’t have full real-time autonomy. But the results mark a clear leap forward in the effort to merge AI reasoning with physical capability.

If you’re following the evolution of robotics, this kind of breakthrough is the story behind the story. AI is quietly learning how to do everything better.

🔗 Read the full article at The Verge