Neural ATTF: A Scalable Solution to Lifelong Multi-Agent Path Planning

TL;DR

Neural ATTF addresses lifelong MAPD by decoupling task assignment from space-time path planning and integrating a learning-guided planner. The PGTM module delivers globally informed task allocation prioritizing delayed agents, while Neural STA* provides fast, collision-free space-time trajectories via a differentiable A* cost map. Across extensive benchmarks against TPTS, CENTRAL, RMCA, LNS-PBS, and LNS-wPBS, Neural ATTF achieves superior scalability and competitive service times with substantially lower runtimes, particularly in large, dynamic settings. The framework also incorporates idling and deadlock recovery and explicit handling of delays, demonstrating robustness and practical viability for real-time multi-agent coordination in unpredictable environments.

Abstract

Multi-Agent Pickup and Delivery (MAPD) is a fundamental problem in robotics, particularly in applications such as warehouse automation and logistics. Existing solutions often face challenges in scalability, adaptability, and efficiency, limiting their applicability in dynamic environments with real-time planning requirements. This paper presents Neural ATTF (Adaptive Task Token Framework), a new algorithm that combines a Priority Guided Task Matching (PGTM) Module with Neural STA* (Space-Time A*), a data-driven path planning method. Neural STA* enhances path planning by enabling rapid exploration of the search space through guided learned heuristics and ensures collision avoidance under dynamic constraints. PGTM prioritizes delayed agents and dynamically assigns tasks by prioritizing agents nearest to these tasks, optimizing both continuity and system throughput. Experimental evaluations against state-of-the-art MAPD algorithms, including TPTS, CENTRAL, RMCA, LNS-PBS, and LNS-wPBS, demonstrate the superior scalability, solution quality, and computational efficiency of Neural ATTF. These results highlight the framework's potential for addressing the critical demands of complex, real-world multi-agent systems operating in high-demand, unpredictable settings.

Figures (5)

• Figure 1: Overall flow of Neural ATTF in each timestep
• Figure 2: Training Pipeline of Neural A*. The environment is encoded to get a guidance map, which is used by differentiable A* module to generate the shortest path and search history. A loss between the search history and the ground-truth path is back-propagated to train the encoder.
• Figure 3: Illustration of Neural ATTF’s space-time path planning using Neural Space-Time A*. It depicts the 3D space-time planning grid, where each agent’s planned trajectory is shown along the time axis. Neural STA* uses dynamic occupancy maps to avoid collisions across time, enabling coordinated, collision-free navigation in multi-agent scenarios.
• Figure 4: These figures represent warehouse environments for the MAPF experiments. (a) is a $63 \times 161$ 4-connected grid with 500 agents, (b) is a $32 \times 32$ 4-connected grid with 100 agents, (c) is a $33 \times 46$ 4-connected grid with 190 agents. Black cells are blocked. Gray cells are task endpoints. Colored circles are the initial locations of agents.
• Figure 5: These figures represent warehouse environments for the MAPD experiments. (a) is a $35 \times 21$ 4-connected grid with 50 agents, (b) is an $101 \times 81$ 4-connected grid with 500 agents. Black cells are blocked. Gray cells are task endpoints. Colored circles are the initial locations of agents.