Probabilistic Recurrent Intention Switching Model

ArXi:2605.26998v1 Announce Type: new Inverse reinforcement learning (IRL) recovers reward functions from observed behavior, yet traditional methods assume a single stationary reward that cannot capture goal switching within an episode. Recent multi-intention IRL methods address this by segmenting trajectories, but model intention transitions as either a memoryless Marko chain or via manual state augmentation with a fixed history window.