Trajectory-Based Active Inverse Reinforcement Learning for Learning from Demonstration

Inverse Reinforcement Learning (IRL) requires a large number of demonstrations for robust learning, especially in the context of Learning from Demonstrations (LfD). Previous Active IRL algorithms have focused on stateaction pair demonstrations and heuristic query strategies. This study introduces Trajectory-based Active Learning for IRL, utilizing acquisition functions derived from Uncertainty sampling in Active Learning. Our novel acquisition function systematically assesses the informativeness of trajectory demonstrations originating from a specific state. Experimental evaluation in the Object World environment shows a significant performance improvement over baselines. By selectively querying informative starting states, we achieve substantial progress in learning from demonstrations.