[Paper Review] Overview: Generalizations of Multi-Agent Path Finding to Real-World Scenarios

Hang Ma, Sven Koenig+8 more|arXiv (Cornell University)|Feb 17, 2017

Robotic Path Planning Algorithms17 references90 citations

TL;DR

This paper surveys how MAPF generalizes to real-world settings and outlines four research directions—TAPF, PERR, motion-predictability via highways, and plan-execution postprocessing—to leverage problem structure and ensure practical execution.

ABSTRACT

Multi-agent path finding (MAPF) is well-studied in artificial intelligence, robotics, theoretical computer science and operations research. We discuss issues that arise when generalizing MAPF methods to real-world scenarios and four research directions that address them. We emphasize the importance of addressing these issues as opposed to developing faster methods for the standard formulation of the MAPF problem.

Motivation & Objective

• Motivate why faster MAPF methods for the standard problem are insufficient for real-world scenarios.
• Introduce four research directions that address real-world structure and execution constraints.
• Showcase how these directions exploit problem structure to improve practicality and robustness.
• Highlight the importance of plan execution considerations alongside path planning.

Proposed method

• Formulate Combined Target Assignment and Path Finding (TAPF) for teams of agents and present an optimal TAPF method scalable to dozens of teams and hundreds of agents.
• Define Package-Exchange Robot Routing (PERR) to allow payload transfers and analyze its complexity, showing MAPF/TAPF relatives with NP-hard approximation results.
• Develop a highway-based scheme to improve path consistency and predictability by biasing MAPF solutions toward user-provided or automatically generated highways.
• Propose a postprocessing framework using a temporal network to convert MAPF outputs into executable plans that respect kinematic constraints and safety distances.

Experimental results

Research questions

• RQ1 How can target assignment be integrated with path finding to optimize makespan for teams of interchangeable agents?
• RQ2 What are the implications and limits of allowing payload transfers in MAPF, and how does this affect complexity and solvability?
• RQ3 Can structured highways improve the consistency and predictability of multi-agent motions in shared workspaces?
• RQ4 How can MAPF solutions be adapted to account for imperfect execution and still avoid replanning?
• RQ5 What is the trade-off between exploiting problem structure and maintaining computational tractability in real-world MAPF variants?

Key findings

• TAPF generalizes standard MAPF and the anonymous MAPF variant, with a Conflict-Based Min-Cost Flow method achieving optimal TAPF scaling to many teams and hundreds of agents.
• PERR extends MAPF to allow package exchanges, proving hardness of approximating optimal solutions and showing that MAPF and TAPF are NP-hard to approximate within 4/3 for makespan with even two teams in TAPF.
• Exploiting problem structure with highways accelerates MAPF solvers while maintaining bounded-suboptimal costs.
• A simple temporal-network postprocessing framework can create executable plan schedules that account for bottlenecks, kinematic constraints, and safety distances, reducing the need for replanning.
• The work emphasizes real-world applicability by linking planning with execution in a cohesive framework.