Researchers have developed a new method for learning symbolic models of real-world robot tasks from visual demonstrations. The PO-PDDL system constructs Partially Observable Markov Decision Process (POMDP) models that can handle both stochastic action execution and partial observability. This is achieved through a demonstration-driven pipeline that reconstructs latent state trajectories, identifies partial observability, and learns transition and observation models. The resulting models are reusable across tasks and enable online belief-space planning under uncertainty. Experiments show that PO-PDDL outperforms existing methods in robust task planning with lower planning cost. --- Why it matters: This matters to engineers because it provides a more efficient way to learn symbolic models of complex robot tasks, which can be used for online planning and decision-making under uncertainty. Source: https://arxiv.org/abs/2606.15654