PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning

Many planning applications involve complex relationships defined on\nhigh-dimensional, continuous variables. For example, robotic manipulation\nrequires planning with kinematic, collision, visibility, and motion constraints\ninvolving robot configurations, object poses, and robot trajectories. These\nconstraints typically require specialized procedures to sample satisfying\nvalues. We extend PDDL to support a generic, declarative specification for\nthese procedures that treats their implementation as black boxes. We provide\ndomain-independent algorithms that reduce PDDLStream problems to a sequence of\nfinite PDDL problems. We also introduce an algorithm that dynamically balances\nexploring new candidate plans and exploiting existing ones. This enables the\nalgorithm to greedily search the space of parameter bindings to more quickly\nsolve tightly-constrained problems as well as locally optimize to produce\nlow-cost solutions. We evaluate our algorithms on three simulated robotic\nplanning domains as well as several real-world robotic tasks.\n