Symbolic Planning

Planning that describes states and actions as logical symbols, then searches for an action sequence that reaches a goal.

Symbolic planning abstracts the world into a set of logical propositions, such as (on cup table) or (handempty), and writes each action as a precondition plus an effect (which facts it adds or deletes). It then searches from the initial state for a sequence of actions that makes all the goal propositions true. The field traces back to STRIPS, developed by Richard Fikes and Nils Nilsson at SRI in 1971 for the Shakey robot; today's common planning description language, PDDL, is built on it, and solving typically uses heuristic-guided state-space search. Classical planning assumes the initial state is known and action outcomes are deterministic, so its results are verifiable and explainable — but turning perceived objects into symbols, and grounding symbolic actions back into continuous motion, is left to task-and-motion planning. The large-language-model era has brought approaches where an LLM translates natural language into PDDL and hands it to a classical planner to solve, such as LLM+P (2023).

ExampleA PDDL ‘pick up’ action might be written with parameters ?o (an object) and ?l (a location); precondition (at robot ?l), (on ?o ?l), (handempty); and effect: add (holding ?o), delete (on ?o ?l) and (handempty). From rules like this, a planner can automatically sequence a plan such as ‘walk to the table → pick up the cup → walk to the sink → put it down.’

Also called

Classical Planning, STRIPS, Automated Planning

Related

Planning Domain Definition Language · Task Planning · Task and Motion Planning · Hierarchical Task Network · LLM-based Task Planning · PDDLStream

Sources

Stanford Research Institute Problem Solver (STRIPS) - Wikipedia
Automated planning and scheduling - Wikipedia（经典规划的假设与求解方法） (Chinese)
LLM+P: Empowering Large Language Models with Optimal Planning Proficiency (arXiv 2304.11477)