Task and Motion Planning

Jointly solving both which actions to take and exactly how each action should move, at the same time.

Task and motion planning (TAMP) decides two kinds of problems together: the discrete task level (which object to move, where to put it, in what order) and the continuous motion level (grasp poses, placement positions, collision-free arm trajectories). A 2021 survey by Garrett, Kaelbling, and Lozano-Pérez et al. notes that TAMP spans discrete task planning, discrete-continuous mathematical programming, and continuous motion planning, and that none of these fields alone solves it well. The reason is that the two levels depend on each other: a plan that is symbolically valid may be physically unreachable or blocked, while pure motion planning can't handle long sequences of decisions. A representative tool, PDDLStream, extends PDDL with ‘streams’ that treat inverse kinematics, grasp sampling, and collision checking as black-box samplers to call during search.

ExampleAsking a robot to put a green block into a cabinet whose door is blocked by a red block: pure task planning doesn't know the red block is in the way, and pure motion planning can't find a feasible path; TAMP produces a plan that first moves the red block aside, while simultaneously computing where to move it, how to grasp it, and how the arm should travel.

Also called

TAMP, Integrated Task and Motion Planning

Related

Task Planning · Motion Planning · Planning Domain Definition Language · PDDLStream · Long-horizon Task · LLM-based Task Planning

Sources

Integrated Task and Motion Planning (Garrett et al., Annual Review of Control, Robotics, and Autonomous Systems 2021)
PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning