PDDLStream

A framework that combines a symbolic planner with continuous samplers to do task and motion planning.

PDDLStream is a planning framework proposed by Caelan Garrett, Tomás Lozano-Pérez, and Leslie Kaelbling at MIT, with the paper published at ICAPS 2020 and the code open-sourced. It extends PDDL (Planning Domain Definition Language, a standard way of describing planning problems) with “streams”: conditional samplers that generate continuous values on demand, such as grasp poses, placement locations, inverse-kinematics solutions, or collision-free paths. The planner first assumes these values exist and optimistically searches out a symbolic plan, then calls the samplers to verify and fill in the actual values, backtracking and re-searching if that fails. It addresses a gap between pure symbolic planners, which can't handle continuous geometry, and pure motion planners, which don't understand task ordering — making it a common baseline in task and motion planning (TAMP).

ExampleHave a robot arm retrieve a cup blocked by a box: PDDLStream plans “move the box aside, then grasp the cup,” and uses its streams to sample a grasp pose and collision-free trajectory for each step.

Related

Task and Motion Planning · Planning Domain Definition Language · Symbolic Planning · Motion Planning · Inverse Kinematics (IK)

Sources

caelan/pddlstream (GitHub)
PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning (arXiv)