Pelican-Sim: New AI Lets Robots Practice in Their Heads, Not Your Kitchen

A research team has published Pelican-Sim 1.0, a simulator that lets robots practice in their imagination instead of in the real world. The key idea is a "world model" — an AI that watches a robot's camera view and its movement commands, then predicts what the robot will see a moment later. Today, teaching a robot to fold laundry or pick up a box means physically running it over and over, which is slow, expensive and breaks things. Pelican-Sim replaces much of that trial and error with fake-but-realistic video, the way pilots log hundreds of hours in a flight simulator before ever taking off.

What makes it unusual is that one model covers many different robots. The team built a universal way to describe movement (a single 28-number recipe that works for arms, grippers, wheeled bases and more), so the same AI isn't rebuilt for every machine. It also uses a "mixture of experts" setup — many small specialist brains, with only the relevant ones switching on for each situation, like a company where you only call the department you need. And it's fast: it generates its predictions in four steps rather than 35, about 5.7 times quicker, which matters when robots must react in real time.

After training on about a million real and simulated robot runs, the results got attention. The predicted video was dramatically sharper and more faithful to the robot's actual commands than competing systems. More importantly, the team showed the simulator can be a teacher: for each task, they added 500 computer-generated practice runs to just 50 real demonstrations, and the robot's success rate jumped from 70% to 93%. The simulator also proved accurate at grading robots, agreeing with real-world testing almost perfectly — a cheap way to check whether a robot has improved without running it.

The catch: it is still pretend. Simulations never capture everything — slippery floors, odd lighting, a cat wandering into frame — and small mismatches between the fake world and the real one can trip robots up. This is a technical report, not a product you can buy, and Pelican-Sim's arrival also speeds up automation of warehouse and factory jobs. Still, the direction is clear: robots that rehearse in software will get good, and cheap, faster than robots that learn by breaking real things.