Hierarchical Task Network

Breaking a large task down, layer by layer, into directly executable actions using human-written ‘decomposition methods.’

The hierarchical task network is a form of classical AI planning, with roots in 1970s planners such as NOAH, formally defined by Erol, Hendler, and Nau in 1994. Tasks are split into two kinds: primitive tasks, which can be executed directly, and compound tasks, which need further decomposition. A domain expert writes ‘methods’ for each compound task, specifying under what conditions it breaks into which subtasks and in what order; the planner then repeatedly decomposes the top-level task until only a sequence of primitive actions remains. Compared to PDDL, which just gives a goal and lets the planner search for a combination of actions on its own, HTN leans on human-authored knowledge to search faster and produce more predictable results, at the cost of someone having to write the methods by hand. A representative system is SHOP2. In robotics it's commonly used for high-level decomposition of long-horizon tasks, and often paired with LLM-based task planning and behavior trees.

Example‘Clear the table’ decomposes into ‘collect the dishes’ and ‘wipe the table’; the method for ‘collect the dishes’ further decomposes into ‘move to the table → identify a bowl → pick up the bowl → put it in the sink,’ repeating while dishes remain — with every bottom-level step corresponding to a skill the robot already has.

Also called

HTN, HTN Planning

Related

Task Planning · Planning Domain Definition Language · Symbolic Planning · LLM-based Task Planning · Behavior Tree · Long-horizon Task

Sources

Erol, Hendler, Nau: HTN Planning: Complexity and Expressivity (AAAI 1994)
Nau et al.: SHOP2: An HTN Planning System (JAIR 2003)
Höller et al.: HDDL: An Extension to PDDL for Expressing Hierarchical Planning Problems (AAAI 2020)