Simultaneous learning of a robot's gait and its body dimensions and compliances

Overview The goal of this project was to have a simulation simultaneously learn a robot's gait as well as body dimensions and compliances—in essence, create a design through machine learning. Can we generate an optimal gait and body design instead of spending much time hand-tuning these parameters? As a practical motivation, the robot used in the simulation was loosely modeled after an actual robot, iSprawl, created by Sangbae Kim in Prof. Mark Cutkosky's lab in the Mechanical Engineering department here at Stanford, though some details of the simulation differed from the actual robot. Figure 1 shows pictures of the simulated and iSprawl robots. Originally, we simulated a quadruped robot with three degrees of freedom per leg, including a knee joint, for a total of 62 parameters; however, this was much too complicated for the simulation to learn good parameters effectively, so we switched to this simpler robot. The simulation was done using the Arachi Dynamics Engine (Arachi, Inc.) and RiSE modeling environment used by the Robotics in Scansorial Environments DARPA project.