Robot Dog Teaches Itself to Walk in Only an Hour

Researchers from the Max Planck Institute for Intelligent Systems in Stuttgart constructed a robot dog and a learning algorithm to study how animals learn to take their first steps. The robot quadruped, named Morti, possessed reflexes similar to those found in an animal and gained the ability of mobility in about one hour. Learning to walk was assisted by a Bayesian optimization algorithm where Morti’s measured foot sensor information matches with target data from a modeled virtual spinal cord acting as a program. Morti then identifies the walking process by modifying motor control patterns, running reflex loops, and comparing expected and sent sensor information. Morti’s learning algorithm alters the control parameters of a central pattern generator (CPG). Animals and humans use CPGs, which are biological neural circuits in the spinal cord that make rhythmic outputs without input from the brain. CPG networks help with rhythmic, involuntary tasks such as walking or blinking. The researchers put Morti’s CPG on a virtual spinal cord that controls the motion of the legs. While Morti figured out how to walk, sensors on its feet are compared to the CPG’s anticipated touch-down. If Morti found a bump in the road, the algorithm would adjust the legs’ speed, time on the ground, and how far they swung back and forth. Felix Ruppert, a research team member, said, “Our robot is ‘born’ knowing nothing about its leg anatomy. If the sensor data doesn’t match the expected data, the learning algorithm changes the behavior until the robot walks … without stumbling.”