Animals and robots use different gaits like walking or trotting for various speeds and terrains. It is not clear which gait is best, especially on rough ground.
Researchers used deep to train a system that executes eleven distinct gaits on challenging terrain, combining it with a model of spinal rhythm generators. A second policy then selects the best gait for the current conditions, optimizing energy, stability, or smoothness.
Every gait is the best choice under some combination of terrain, speed, and goal. This shows that the variety of gaits is functional. For robots, picking the right gait can improve performance.
