SMART: Scalable Multi-Agent Realistic Testbed

Advancing MAPF Toward the Real World: A Scalable Multi-Agent Realistic Testbed
(SMART)

Jingtian Yan¹, Zhifei Li¹, William Kang¹, Kevin Zheng¹, Yulun Zhang¹,
Zhe Chen¹, Yue Zhang¹, Daniel Harabor², Stephen F. Smith¹, Jiaoyang Li¹

¹Carnegie Mellon University    ²Monash University

IEEE Robotics and Automation Letters, vol. 11, no. 6, pp. 7428-7435, 2026

▤PaperIEEE Xplore ⌁PreprintarXiv ⌘CodeGitHub ▶Live DemoTry SMART

★

ICAPS 2025 Best Demo AwardPresented at the International Conference on Automated Planning and Scheduling

INTERACTIVE DEMO

Use our online interface

Run SMART directly in your browser—no installation required.

smart-mapf.github.io/demoLaunch interface ↗

DEMO VIDEO

SMART in action

OVERVIEW

Closing the simulation-to-reality gap

State-of-the-art MAPF algorithms plan paths for hundreds of robots in seconds, but their simplifying assumptions rarely survive real-world execution.

They often use simplified robot models that ignore kinodynamic constraints and assume robots execute paths perfectly. SMART fills this gap with physics-based simulation, realistic robot behavior, and robust execution monitoring.

• Realistic environmentsPhysics engines model kinodynamics and execution uncertainty.
• Flexible executionAn Action Dependency Graph monitor supports different MAPF algorithms and robot models.
• Massive scalabilityEfficient simulation enables experiments with thousands of robots.

KEY FEATURES

Built for realistic MAPF evaluation

A testbed designed to expose the challenges that matter in deployment.

01

Robust execution

Account for delays and uncertainty while preserving collision-free execution through ADG monitoring.

02

Scalable experiments

Move from small debugging scenarios to stress tests involving thousands of coordinated agents.

03

Broad planner & map support

Evaluate different MAPF planners across benchmark, warehouse, and custom maps through a consistent simulation interface.

RESULTS GALLERY

Across simulators and robots

High-fidelity simulationKinodynamic execution in NVIDIA Isaac Sim.

Physical deploymentFrom planned paths to real multi-robot execution.

CITE THIS WORK

Publication

Yan et al., “Advancing MAPF Toward the Real World: A Scalable Multi-Agent Realistic Testbed (SMART),” IEEE Robotics and Automation Letters, 2026.

BibTeX

@article{yan2026smart,
  title={Advancing MAPF Toward the Real World: A Scalable Multi-Agent Realistic Testbed (SMART)},
  author={Yan, Jingtian and Li, Zhifei and Kang, William and Zheng, Kevin and Zhang, Yulun and Chen, Zhe and Zhang, Yue and Harabor, Daniel and Smith, Stephen F. and Li, Jiaoyang},
  journal={IEEE Robotics and Automation Letters},
  volume={11}, number={6}, pages={7428--7435},
  year={2026}, month={June}, doi={10.1109/LRA.2026.3688062}
}