LSMART

Target audience and objectives

1. For multi-robot researchers and practitioners, this tutorial targets both (i) those who are already familiar with MAPF but may view it as primarily theoretical and limited to grid-based, discrete-time settings, and (ii) those who are new to MAPF and are seeking scalable and practical tools for coordinating large robot teams. The tutorial will present MAPF as a powerful and scalable algorithmic foundation for coordinating large robot teams and demonstrate how modern MAPF-based methods can be extended to handle robot dynamics, execution uncertainty, and real-world constraints, and how they can be deployed on various practical robotic platforms like ground robots, robot arms, and aerial robots.

2. For AI and MAPF researchers, this tutorial will highlight the importance of real-robot deployment for MAPF research and present key open challenges and research opportunities that arise when bringing MAPF algorithms to real robots.

3. For both communities, this tutorial will provide an end-to-end view of multi-robot coordination—from planning and execution to learning and deployment—and clarify how recent MAPF advances translate into scalable, high-quality solutions for real robotic systems.

4. For students and developers, this tutorial will offer hands-on experience with state-of-the-art MAPF-based tools, enabling participants to directly experiment with scalable multi-robot planning and execution pipelines.

Real-world Considerations

Built on top of SMART [1], we consider realistic AGV kinodynamics by modeling them as differentiable drive robots, which can move forward and rotate in place with constraints in maximum velocity and acceleration. This is in contrast to prior works that model AGVs as simple omnidirectional/pebble motion agents. We also simulate real-world execution delays and use Action Dependency Graph (ADG) [2] to ensure collision-free of the executed paths.

Insane Scalability

LSMART is highly scalable in terms of the number of robots and the size of the map. Here we show a simulation of 1000 robots in the warehouse-10-20-10-2-1 map from the MAPF benchmark [3]. We use PIBT [4] as the underlying MAPF planner and plan windowed paths every 1 simulation second. The simulation runs for 600 simulation seconds and takes 268 seconds to finish with a AMD Ryzen 9 9950X 16-Core Processor.

Preparation

Participants must bring a laptop (macOS, windows, or Linux). Installing Docker in advance is strongly recommended for the hands-on exercises.

Prerequisites

Participants should understand the fundamentals of Multi-Agent Path Finding (MAPF) and have basic skill on running commands in terminals.

References

[1] Yan, J.; Li, Z.; Kang, W.; Zheng, K.; Zhang, Y.; Chen, Z.; Zhang, Y.; Harabor, D.; Smith, S. F.; and Li, J. 2025. Advancing MAPF towards the Real World: A Scalable Multi-Agent Realistic Testbed (SMART). ArXiv, abs/2503.04798.

[2] Hönig, W.; Kiesel, S.; Tinka, A.; Durham, J. W.; and Ayanian, N. 2019. Persistent and Robust Execution of MAPF Schedules in Warehouses. IEEE Robotics and Automation Letters, 4: 1125-1131.

[3] Stern, R.; Sturtevant, N. R.; Felner, A.; Koenig, S.; Ma, H.; Walker, T. T.; Li, J.; Atzmon, D.; Cohen, L.; Kumar, T. K. S.; Barták, R.; and Boyarski, E. 2019. Multi-Agent Pathfinding: Definitions, Variants, and Benchmarks. In Proceedings of the International Symposium on Combinatorial Search (SoCS), 151-159.

[4] Okumura, K.; Machida, M.; Défago, X.; and Tamura, Y. 2019. Priority Inheritance with Backtracking for Iterative Multi-agent Path Finding. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 535-542.