Robots Learn New Tasks Without Forgetting Old Ones

Imagine teaching your robot vacuum a new trick—like sorting recycling—and it suddenly forgets how to vacuum. That's a real problem called catastrophic forgetting, and it's a major roadblock for robots that need to keep learning new tasks throughout their lives. Most robots try to avoid this by replaying old training memories while learning new ones, like a student reviewing old flashcards. But the way they choose those memories is usually random.

A team of researchers from Stanford and Columbia discovered something surprising: not all memories are equal. They found a small set of "memory anchors"—the moments where a new task conflicts with an old one—that are critical for remembering past skills. It's like when you learn to use a familiar tool in a new way; the confusion point is exactly what your brain uses to keep the two methods separate. If you remove just 10% of these anchors, forgetting increases by more than 4.5 times. But if you deliberately include them, forgetting drops by 63%.

What's even more impressive is that this works on real robots, not just simulations. A physical robot successfully learned two tasks back-to-back without forgetting the first one. That's a big step from lab-only breakthroughs.

Why should you care? Because this makes future robots cheaper and more practical. Instead of sending a robot back to the factory to be reprogrammed every time it needs a new skill, it can learn on the job. That means home robots that adapt to your routines, warehouse robots that take on new tasks, and less downtime overall.