Planning Domain Definition Language

规划领域定义语言PDDLAdvanced

A standard language for writing down actions' preconditions, effects, and a task's goal, for a general-purpose planner to solve.

PDDL is the standard description language of classical AI planning, created in 1998 by Drew McDermott and colleagues for the International Planning Competition (IPC), drawing on earlier planning formalisms such as STRIPS and ADL. It splits a planning problem into two parts: a domain file, which defines predicates (true/false statements describing world state, such as ‘block A is on block B’) and actions (each action's preconditions and the effects of executing it); and a problem file, which gives the specific objects, initial state, and goal. Once written, this is handed to a general-purpose planner to automatically search out a sequence of actions. Later versions added numeric and durative actions (PDDL2.1), continuous processes and events (PDDL+), and preferences and trajectory constraints (PDDL3.0). In robotics it commonly serves as the symbolic interface at the task-planning layer: the task-and-motion-planning framework PDDLStream wires continuous-parameter samplers — for grasp poses, placement positions, and so on — into PDDL as ‘streams’; LLM+P has a large language model translate a natural-language instruction into PDDL, then hands it to a classical planner to compute a correct plan, addressing how easily an LLM goes wrong doing long-horizon planning directly.

Example‘Put the cup on the table into the cabinet’ can be written in PDDL as: action pick(?obj) has precondition ‘hand empty’ and ‘object reachable,’ with effect ‘holding the object’; further defining open(?door) and place(?obj ?loc), and giving the initial state and the goal ‘cup is in the cabinet,’ a planner outputs ‘open the cabinet door → pick up the cup → place it in the cabinet.’

Also called: PDDL
Related: Task Planning · Symbolic Planning · Task and Motion Planning · PDDLStream · LLM-based Task Planning · Hierarchical Task Network
Sources: Wikipedia: Planning Domain Definition Language
LLM+P: Empowering Large Language Models with Optimal Planning Proficiency (arXiv 2304.11477)
PDDLStream: Integrating Symbolic Planners and Blackbox Samplers (arXiv 1802.08705)