Machine learning for robot control learns to doubt itself — and uses fewer training examples

Researchers built a system that teaches robots to move and balance using fewer trial-and-error attempts, by having the AI estimate how confident it should be in its own predictions. Instead of averaging over all possible movements and getting a blurry answer, the system tracks multiple possibilities at once and discounts the predictions it's least sure about during training.

Model-based reinforcement learning has always promised to be data-efficient — the robot learns by building a mental model of physics, then testing it, rather than brute-forcing millions of random tries. But in practice it fails because the model's errors compound: a small mistake in predicting what happens next becomes a bigger mistake in the prediction after that, and the robot learns from corrupted data. This paper shows a structural fix: weight the training examples by how confident the model is in them. On a hard benchmark task (a humanoid robot learning to run), the system cut the number of training examples needed by more than half. The benchmark numbers matter because sample efficiency is what would make this technology useful in the real world — fewer training attempts means fewer expensive robot hours.