Robots Atlas>ROBOTS ATLAS
Robotics

Task Planning

1971ActivePublished: 21 October 2026Updated: 21 October 2026Published
Key innovation
Separating the decision of "what to do and in what order to reach a goal" from low-level control, via an explicit symbolic representation of state, actions and goal over which a plan can be searched automatically.
Category
Robotics
Abstraction level
Pattern
Operation level
Robot controlAgent runtimeOrchestration
Use cases
Task planning for humanoid and manipulation robotsTask-and-motion planning (TAMP)Autonomous agents and LLM tool orchestrationMobile robotics and high-level navigationLogistics and operations schedulingGames and symbolic problem solving

How it works

1. Representation: the world state is described by a set of predicates/facts, and actions as operators with preconditions and effects (added and deleted facts). 2. Goal specification: the goal is a partial description of the desired state. 3. Search: the planner searches the state space (forward, backward, or in plan space), often with heuristics, to find a sequence of actions leading from the initial state to the goal. 4. Decomposition (HTN): compound tasks are recursively decomposed into subtasks down to primitive actions. 5. Motion integration (TAMP): the task plan is checked for geometric feasibility by a motion planner; infeasible steps trigger backtracking and replanning. 6. LLM variant: a large language model generates candidate steps in natural language, and a grounding mechanism (e.g. skill value functions) selects actions that are actually feasible for the given robot.

Problem solved

How to automatically translate a high-level goal ("make coffee", "clear the table") into a concrete, executable sequence of actions, accounting for preconditions, effects and environment constraints โ€” without hand-coding every situation. Task planning handles the decision layer of "what to do", decoupled from "how to physically execute it".

Components

State representationDescribes the world the planner reasons over.

A set of predicates/facts describing the current world state (e.g. at(robot, kitchen), holding(cup)).

Official

Action operatorsDefine which state changes are possible.

Action schemas with preconditions and effects (added/deleted facts), e.g. in STRIPS/PDDL.

Official

Goal specificationDefines the success criterion for planning.

A partial description of the desired final state the plan must achieve.

Planner (search algorithm)Produces a plan from the domain definition and goal.

The engine that searches the state or plan space (e.g. with heuristics) to find a valid plan. Variants: classical, HTN, and LLM-based planners.

Classical planning (STRIPS/PDDL)State-space search with heuristics.
HTNHierarchical task decomposition.
LLM-based planningPlan steps generated by a language model.

Official

PlanThe planning output passed to execution.

The resulting ordered sequence (or partial order / policy) of actions leading to the goal.

Evolution

Original paper ยท 1971 ยท Artificial Intelligence, 1971 (SRI International) ยท Richard E. Fikes
STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving
Richard E. Fikes, Nils J. Nilsson
1971
STRIPS โ€” the first influential planner and action language
Inflection point

Fikes and Nilsson at SRI International define STRIPS; its language became the basis for most modern planning languages.

1998
PDDL โ€” standardization of the planning language
Inflection point

Drew McDermott and colleagues introduce PDDL for the International Planning Competition, standardizing planning domain descriptions.

PDDL โ€” The Planning Domain Definition Language (paper)
2020
Integrated Task and Motion Planning (TAMP) โ€” survey

Garrett et al. formalize and organize methods for integrating task planning with motion planning in robotics.

Integrated Task and Motion Planning (paper)
2022
SayCan โ€” LLM-based task planning for robots
Inflection point

Ahn et al. (Google) combine LLM knowledge with skill value functions to produce feasible high-level plans for robots.

Do As I Can, Not As I Say: Grounding Language in Robotic Affordances (paper)

Hyperparameters (configurable axes)

Planning horizonHigh

How far into the future (how many steps) the planner reasons.

Search strategyHigh

Search method: forward / backward / plan-space, with or without heuristics.

Action granularityMedium

The abstraction level of operators โ€” from coarse tasks to fine primitive actions.

Computational complexity

Time complexity: PSPACE-complete.

Hardware requirements

Primary

Classical and hierarchical planning is symbolic search โ€” primarily a CPU operation, independent of accelerators.

Possible

LLM-based variants require a GPU for inference of the model that generates plan steps.