In this paper, we present a novel framework, Quality Diversity Inverse Reinforcement Learning (QD-IRL), to overcome the limitations of single-expert policy learning and enable diverse and robust robot locomotion. We integrate quality diversity optimization with IRL techniques to learn diverse behaviors from limited demonstration data. In particular, we improve the exploration of diverse walking behaviors by introducing Extrinsic Behavioral Curiosity (EBC), which receives additional curiosity rewards based on novelty of the behavioral archive from external evaluators. We evaluate EBC along with GAIL, VAIL, and DiffAIL on several robot locomotion tasks and demonstrate its performance improvement, outperforming expert performance by up to 20% in a humanoid environment. In addition, we show that EBC can be applied to Gradient-Arborescence-based QD-RL algorithms.
