Humanoid robots need to learn how to track human-like motions quickly. Traditional training methods treat the simulator as a black box and require millions of samples, which takes hours on powerful computers. Making the training faster is important because it would let robots learn from thousands of different movements rather than just a few.
This research introduces a technique that uses exact gradients from the simulation engine instead of learning a separate model. The authors solved two problems: extreme gradient values that can break the training, and the high computational cost of calculating gradients. They clip extreme gradients and reuse simulation data more efficiently.
The new method trains a single policy to track all 40 movements from the LAFAN1 dataset. It reaches the score of the state-of-the-art approach in 20 minutes, about four times faster. This speedup makes it practical to train on much larger motion datasets, which could lead to more capable humanoid robots.