Robots that walk on rough ground usually learn on a ladder of manufactured obstacle courses. A new method throws out the ladder and lets the robot build its own.
Researchers describe a system called TACL - Trajectory-level Automatic Curriculum Learning - that trains legged robots directly on unstructured terrain maps instead of hand-designed terrain parks. The core problem: real-world terrain has no built-in difficulty ranking, so prior methods relied on engineers manually staging harder and harder synthetic obstacles. That works, but robots trained this way tend to memorize the shape of those specific obstacle patterns rather than learning to handle genuinely unpredictable ground. TACL instead uses an evaluator that scores how hard a given trajectory is for the robot's current skill level, then a sampler proposes new practice routes based on that score, adjusting as the robot improves.
The results matter because terrain generalization has been a persistent bottleneck in legged robotics, not a solved problem being polished. The paper reports a 56.3 percent improvement in trajectory success rate over training with no curriculum at all, and an 18.5 percent improvement over handcrafted curricula on the hardest terrain tasks. On the same obstacle approached from different directions, the gain reached nearly 40 percent - a sign the robot is learning the terrain itself, not a fixed script for crossing it.
Handcrafted curricula have been the default in this field for years precisely because they are easy to control and debug. An automatic, closed-loop alternative that outperforms them on directional variation suggests the field's dependence on human-tuned obstacle courses may be more of a crutch than a necessity.