PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty
Researchers have developed PO-PDDL, a new symbolic formulation for Partially Observable Markov Decision Processes (POMDPs) that maintains the relational structure of PDDL while explicitly modeling partial observability and stochasticity. This formulation is designed to be LLM-friendly and reusable across tasks. A pipeline is proposed to learn PO-PDDL models from visual demonstrations of robot executions, identifying partial observability and learning stochastic models. Experiments on real-world manipulation tasks show that this method outperforms existing PDDL and POMDP learning approaches in robust task planning under uncertainty with reduced planning costs. AI