Most humanoid robots learn to walk by tracking how fast they are moving, but that needs external cameras or complex estimates. STRIDE instead uses only the robot's built-in motion sensor to follow an acceleration command. This makes learning simpler and faster for real robots.
STRIDE trains the robot with a algorithm, which is a trial-and-error method where the robot gets a reward for good behavior. It also uses high-gain joint controllers and a symmetry trick to make learning more efficient and reduce falls.
In simulation, STRIDE gets a humanoid walking in 14 minutes of training data, and it cuts falls during training by 81% compared to not using these tricks. This could speed up real-world robot learning, making it more practical.