IPAB Workshop-02/07/2020

Title: Learning natural locomotion behaviours for humanoid robots using human bias Abstract: In this talk, I will present a learning framework that bridges  knowledge from imitation learning, deep reinforcement learning, and  control theories to achieve human-style locomotion of humanoid robots  that is natural, dynamic, and robust. I proposed novel approaches to  introduce human bias, ie motion capture data and a special  Multi-Expert network structure. I used the Multi-Expert network  structure to smoothly blend behavioral features, and used the  augmented reward design for the task and imitation rewards. The reward  design is more composable, tunable, and explainable by using  fundamental concepts from conventional humanoid control. I rigorously  tested and benchmarked the learning framework on different disturbance  scenarios and consistently produced robust locomotion behaviors.  Further, I demonstrated the capability of learning robust and  versatile policies in presence of disturbances, such as terrain