curriculum learning
Curriculum learning is the practice of teaching a robot the way a good school teaches a child: start with easy tasks, then make them harder step by step, instead of throwing the hardest challenge at it on day one. Just as you would not drop a beginner swimmer into the deep end, you would not ask a freshly initialized robot to sprint over rocky ground; you let it first learn to stand, then to step, then to walk on flat floor, and only later to tackle slopes and rubble. Each easy skill becomes a stepping stone the robot stands on to reach the next, harder one.
It matters because many robot skills are almost impossible to learn from scratch by random trial. If a legged robot starts on brutal terrain knowing nothing, it falls instantly, every single time, and almost never stumbles into the rare lucky sequence of moves that would teach it anything — so learning stalls. By beginning on gentle tasks where success happens often enough to give useful feedback, and then gradually raising the difficulty as competence grows, the robot keeps getting a steady stream of lessons it can actually act on. The ordering of experiences, not just the experiences themselves, is what makes the hard final skill reachable.
Designers build the curriculum in different ways: hand-crafting a fixed sequence of ever-tougher levels, or letting a program automatically nudge the difficulty up whenever the robot is doing well and ease it back down when the robot struggles, so the challenge always sits just past what it can comfortably do. The payoff is faster, more reliable learning and final skills that pure brute-force training often cannot reach at all.
To teach a four-legged robot to cross rough ground, trainers first have it walk on flat virtual floor, then on gentle bumps, then on stairs, and finally on randomly scattered boulders. A robot started straight on the boulders never learns to walk at all; the one led up the curriculum eventually crosses terrain it could never have mastered cold.
Walking is learned in graded stages — flat, bumpy, stairs, boulders — each one preparing the robot for the next.
The name comes by analogy to a school curriculum — a planned order of lessons from simple to advanced — and the idea is most often paired with reinforcement learning, where the difficulty of the practice environment is what gets ramped up.