PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty

The article introduces PO-PDDL, a symbolic framework for learning Partially Observable Markov Decision Processes (POMDPs) that integrates the relational structure of Planning Domain Definition Language (PDDL) while addressing partial observability and stochasticity. The proposed method utilizes visual demonstrations to reconstruct symbolic state trajectories and learn transition and observation models, leading to reusable PO-PDDL domains that facilitate online belief-space planning. Experimental results on long-horizon manipulation tasks demonstrate superior performance compared to existing PDDL and POMDP approaches, highlighting its potential for effective robot planning in uncertain environments.