Task Planning

Given the current state and a goal, finding a sequence of high-level action steps that achieves it.

Task planning is the application of automated planning from AI to robotics: given an initial state, a goal, and a set of available actions (each with preconditions and effects), find a sequence of actions that, once executed, makes the goal true. Classical approaches describe the problem in STRIPS or PDDL (Planning Domain Definition Language) and hand it to a general-purpose planner to search; hierarchical task networks (HTN) instead break a large task down into subtasks layer by layer. Task planning only decides what to do and in what order — not exactly how the arm should move, which is left to motion planning; combining the two gives task and motion planning. Large language models are now commonly used for task planning — SayCan, for instance, has a language model propose candidate skills, then uses each skill's value function to judge whether it can actually succeed in the current scene, grounding the model's language knowledge in a real robot.

ExampleGiven the instruction ‘put the coke on the table into the fridge,’ task planning produces: walk to the table → pick up the coke → walk to the fridge → open the fridge door → put in the coke → close the door, with each step then handed to lower-level skills such as navigation and grasping.

Also called

Automated Planning, AI Planning, High-Level Planning

Related

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

Sources

Wikipedia: Automated planning and scheduling
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances (SayCan)