More about HKUST
Solving Scalability and Reality Gaps in Multi-Agent Pathfinding
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
PhD Thesis Defence
Title: "Solving Scalability and Reality Gaps in Multi-Agent
Pathfinding"
By
Mr. Mingkai TANG
Abstract:
The classical Multi-Agent Pathfinding (MAPF) problem, which involves computing
collision-free paths for multiple agents on a graph, serves as a fundamental
framework for coordinated multi-robot motion. Due to its essential role in
applications such as automated warehousing and logistics, MAPF has
consistently drawn research interest from the fields of artificial
intelligence, multi-agent systems, and robotics. However, its practical
deployment is hindered by two major challenges: the reality gap caused by its
oversimplified problem formulation, and the scalability issue in coordinating
large numbers of agents in complex, long-horizon scenarios.
To bridge the reality gap and enhance the scalability of classical MAPF, this
thesis investigates three key limitations that originate from its simplifying
assumptions: (1) optimizing the order in which agents visit multiple goals,
(2) enabling cyclic operations over an infinite horizon, and (3) accounting
for energy constraints that affect agent viability over time.
First, to enable optimal multi-goal visitation, we introduce the Multi-Goal
Conflict-Based Search (MGCBS) algorithm, whose efficiency is enhanced by a
Time-Interval-Space Forest (TIS Forest) data structure. Second, to accommodate
cyclic, infinite-horizon operations, we formalize the Streaming MAPF (S-MAPF)
problem and propose the Agent Stream Conflict-Based Search (ASCBS) algorithm,
which employs a disjoint-splitting technique to improve efficiency in small
maps. Recognizing that S-MAPF assumes agents disappear upon reaching their
destinations and operate with fixed initial start times, we further extend
this formulation to the Multi-Agent-Stream Scheduling and Pathfinding
(S2-MAPF) problem, which supports persistent agent presence and adaptive
initial start times. For this enhanced model, we introduce two solution
approaches: the Matching Planning (MP) algorithm, which decouples task
scheduling from pathfinding to achieve computational efficiency, and the
Coupled Planning (CP) algorithm, which integrates both subproblems to
guarantee optimality.
Finally, to incorporate energy constraint, we define the Energy-Limited
Lifelong Multi-Agent Pickup and Delivery (EL-MAPD) problem, establish
theoretical solvability conditions, and introduce a practical online planner
called Fallback Priority Planning (FPP), which guarantees a feasible action in
every planning episode.
Together, these contributions establish algorithmic and theoretical
foundations that meaningfully enhance the scalability and real-world
applicability of multi-robot systems.
Date: Monday, 31 August 2026
Time: 10:00am - 12:00noon
Venue: Room 3494
Lifts 25/26
Chairman: Dr. Yantao YU (CIVL)
Committee Members: Dr. Ling PAN (Supervisor)
Dr. Long CHEN
Dr. Qifeng CHEN
Prof. Wei ZHANG (ECE)
Prof. Dongning LIU (Guangdong University of
Technology)