Asynchronous Decentralized Algorithm for Space-Time Cooperative Pathfinding

Cooperative pathfinding is a multi-agent path planning problem where a group of vehicles searches for a corresponding set of non-conflicting space-time trajectories. Many of the practical methods for centralized solving of cooperative pathfinding problems are based on the prioritized planning strategy. However, in some domains (e.g., multi-robot teams of unmanned aerial vehicles, autonomous underwater vehicles, or unmanned ground vehicles) a decentralized approach may be more desirable than a centralized one due to communication limitations imposed by the domain and/or privacy concerns. In this paper we present an asynchronous decentralized variant of prioritized planning ADPP and its interruptible version IADPP. The algorithm exploits the inherent parallelism of distributed systems and allows for a speed up of the computation process. Unlike the synchronized planning approaches, the algorithm allows an agent to react to updates about other agents’ paths immediately and invoke its local spatio-temporal path planner to find the best trajectory, as response to the other agents’ choices. We provide a proof of correctness of the algorithms and experimentally evaluate them on synthetic domains.

💡 Research Summary

**
Cooperative pathfinding concerns the simultaneous planning of space‑time trajectories for multiple agents such that no two agents occupy the same location at the same time. The dominant centralized solution, Prioritized Planning, assigns a static priority order to agents and computes each agent’s path sequentially while treating higher‑priority agents as moving obstacles. Although simple and often fast, this approach requires a central coordinator that collects all agents’ state information, distributes plans, and synchronizes updates. In domains where communication bandwidth is limited, latency is high, or privacy constraints prohibit sharing full trajectories (e.g., swarms of UAVs, AUVs, or ground robots operating in contested environments), a decentralized alternative is highly desirable.

The paper introduces two novel algorithms that bring the prioritized planning paradigm into a fully asynchronous, distributed setting: Asynchronous Decentralized Prioritized Planning (ADPP) and its interruptible variant (IADPP). The key idea is that each agent runs its own local spatio‑temporal planner and maintains a copy of the paths of all higher‑priority agents. Whenever an agent receives a new path from a higher‑priority neighbor, it immediately incorporates the updated constraints and re‑invokes its planner. Because there is no global synchronization barrier, agents can react to updates as soon as they arrive, exploiting the inherent parallelism of a distributed system and tolerating variable network delays.

IADPP extends ADPP by allowing the local planner to be interrupted mid‑search. When a new higher‑priority path arrives while the planner is still exploring the search space, the algorithm saves the current search frontier, aborts the ongoing expansion, and restarts the search with the new constraints while re‑using the previously explored nodes. This mechanism prevents the planner from being stuck in a long, fruitless search when the environment changes frequently, and it reduces the overall response time in densely constrained scenarios.

The authors provide rigorous proofs of correctness for both algorithms. The convergence proof shows that, given a fixed priority ordering, each agent’s set of constraints is monotonically non‑increasing: a newly computed path never violates constraints that were already satisfied. Consequently, each agent can only improve its path a finite number of times before reaching a fixed point, guaranteeing termination. For IADPP, the proof demonstrates that the interrupt‑and‑resume operation preserves the completeness of the underlying planner (e.g., A*), because the saved frontier contains all nodes that were generated but not yet expanded, and these nodes are re‑examined after the restart.

Experimental evaluation is carried out on two synthetic domains. The first is a 2‑D grid world where the number of agents varies from 10 to 100. Four methods are compared: centralized Prioritized Planning, a synchronized distributed version, ADPP, and IADPP. Results indicate that ADPP reduces total planning time by roughly 30 %–60 % compared with the synchronized approaches, and IADPP achieves additional speed‑ups in highly cluttered grids by avoiding costly replanning from scratch. The second domain is a 3‑D aerial simulation representing a swarm of UAVs. The authors inject artificial network latency and packet loss to emulate realistic communication constraints. Even under these adverse conditions, both ADPP and IADPP converge reliably, and the collision‑free success rate remains near 100 % for all methods, confirming the theoretical guarantee of safety.

Despite these strengths, the paper acknowledges several limitations. The priority ordering is static and must be chosen a priori; poor priority choices can degrade performance, and the current framework lacks a mechanism for dynamic priority negotiation. Moreover, the experiments are confined to synthetic benchmarks; real‑world factors such as sensor noise, dynamic obstacles, and heterogeneous communication ranges have not been tested. Future work is suggested in three directions: (1) developing adaptive priority assignment or negotiation protocols, (2) extending the algorithms to handle dynamic environments where obstacles appear or disappear during execution, and (3) deploying the methods on physical robot platforms to assess robustness under real communication constraints.

In summary, the paper makes a significant contribution by translating the well‑understood centralized prioritized planning technique into an asynchronous, decentralized form that is both provably correct and empirically faster in multi‑agent settings with limited communication. The interruptible variant further enhances responsiveness, making the approach promising for real‑time applications in aerial, underwater, and ground robot swarms where centralized control is impractical.