Evaluating symmetry variations in quadruped trot gait locomotion controllers

We compare the performance of eight controllers for quadruped robot trot gaits. The controller variations couple the robot’s actuators using a different symmetry each, to reduce the effective number of parameters required to control the trot gait. We compare the performance of all controllers on a learning task where robots learn to walk fast without falling, in a physically realistic simulation. The parameter spaces of the controllers compared on this task range in dimensionality from seven to nineteen. Successful trot gaits are learned using all controllers tested. We compare the eight different controllers on measures such as stability (falling rate), learning rate and maximum reward achieved under equal but limited computing resources. The controller comparison methods we propose are a step toward automatically choosing among drastically different robotic controllers in general, in context of a specific learned task.