DM-IRL

Inverse Reinforcement Learning (IRL) subcategory of robotic learning from demonstration. This paradigm allows operators to program robots via demonstration instead of explicit programming.

We propose the use of Distance Minimization-Inverse Reinforcement Learning (DM-IRL) as a general purpose IRL method. DM-IRL uses an expert judge to assign scores to demonstrations (trajectories) and removes all optimality requirements from the demonstrations. We show that DM-IRL can learn high-quality behavior from extremely sub-optimal demonstrations sourced from multiple demonstrators with unknown transition functions.

Please see the full paper for more information.