In recent years, the field of embodied intelligence has witnessed significant advances in applications ranging from autonomous robotics and smart transport systems to interactive gaming and virtual agents. These systems rely on the ability to perceive, rational, and act within complex environments. It often includes multiple agents who must coordinate their actions to achieve shared goals. As the complexity and scale of these applications increase, the need for efficient tuning mechanisms becomes most important. One of the key challenges in this domain is the Multi-Agent Pathfinding (MAPF) issue, which plays a key role in enabling multiple agents to navigate the environment without conflicting while optimizing paths. MAPF problem solvers are widely used in a variety of applications, including aircraft towing vehicles, traffic management, email sorting, video games, and more.
In fact, calculation time is the most important bottleneck in solving MAPF problems. The optimal solution to the MAPF problem has proven to be NP hard in many settings, including general graphs, planar graphs, and 2D grid graphs, meaning that it grows exponentially with agents with increasing computational times. In some real applications, excessively long computation times are not acceptable. For example, if the calculation time for the routes of a delivery robot in a warehouse is too long, the throughput rate of the logistics system will be reduced, resulting in economic losses.
Conflict-based Search (CBS) is a common two-level solver for addressing MAPF issues. The high-level solver constructs a constraint tree (CT) that assigns the temporal-spatial constraints for each agent to avoid collisions, while the low-level solver calculates the path under these constraints, including both vertex constraints and edge constraints. Previous art allows for efficient CBS, primarily by reducing the number of nodes investigated by high or low level solvers. Beyond the above methods, GPU-Acceleration has become a promising approach in recent years to improve the efficiency of time-consuming algorithms. However, there is no study investigating GPU acceleration exclusively for CBS computing. This paper takes a new perspective to improve the computational efficiency of CBS by taking significant advantage of the parallel computing capabilities of GPUs.
Deploying the GPU acceleration of CBS algorithms presents several challenges. First, the CBS algorithm is a two-level structure with a highly combined combination, complicating the decomposition of the computational process into components with less data dependencies. This decomposition is important for effective parallel computation. Second, synchronous operations between the various components must be lightweight. This should be lightweight as it can significantly limit the parallel functionality of a GPU. Third, a simultaneous pass funding solver for single agents must be developed within the CBS algorithm. This solver should work in conjunction with a high-level solver, considering the relevant constraints.
In this work published in Machine Intelligence Research, researchers propose a competitive-based search, or GACB, that accelerates the GPU, to significantly improve the computational performance of CBS by efficiently utilizing GPU parallelism. GACBS utilizes the Task Coordination Framework (TCF) to promote lightweight synchronization operations and collaboration between two levels of solvers based on the observation that once constraints are determined, different CT node extensions are independent. For GACBS low-level solvers, a GPU-affiliated time-space A* (GATSA) algorithm has been proposed, and it has been proposed to simultaneously calculate the optimal paths of individual agents under constraints. Additionally, researchers conduct experimental evaluations to demonstrate the effectiveness and efficiency of the GACBS algorithm. Experimental results show that GACBs are significantly more than CPU-based CBS at a maximum speedup ratio of over 46.
Section 2 concerns related work. We first review the CBS development and classify it into nodes where the variant reduces the nodes that were investigated by high-level solvers and low-level solvers. Next, we also point out GPU acceleration in pass funding tasks.
In Section 3, we formulate the MAPF problem for graphs using N-agents, each with a start and target node. At every step, all agents have the option to stay on the current node or move to one of their adjacent nodes. We aim to define a set of conflict-free paths for each agent, minimizing the total cost of all paths.
Section 4 is a preliminary thing. Presents a two-level solver for MAPF, a conflict-based search (CBS). It consists of a high-level solver and a low-level solver. High-level solvers are responsible for building CTs that ensure conflict avoidance between different agents. The low-level solver is responsible for calculating the agent's path while adhering to specified constraints. Within the CT, each node maintains information about the constraint set, the path of each agent, and the total cost.
Section 5 presents GPU-accelerated contention-based search (GACBS), a parallel processing algorithm that utilizes a two-level solver to significantly improve CBS efficiency through GPU acceleration. GACBS integrates three core components: TCF, high-level solver, and low-level solver. TCF promotes inter-solver communication, and high-level solvers build CTs and generate task lists. In particular, the proposed parallel GATSA algorithm serves as a low-level solver that determines the optimal path of the agent based on assigned tasks. This framework efficiently tackles the challenges of the MAPF problem with GPU acceleration.
Section 6 presents a theoretical analysis starting from Lemma 1, which shows that the optimal solution is found under certain conditions when heuristic features are tolerated and consistent. Theorem 1 proves that GATSA is optimal and perfect under such a heuristic. Theorem 2 shows that GACB is also optimal, acceptable and has a consistent, low-level heuristic.
In Section 7, the researcher empirically examines the performance of GACBs on a 4-neighbor grid map. The experiment employs five algorithms: CPUCBS, MIXCBS, GACBS, GACBS-NC, and NRFSAT. Researchers use success rates and successful average running time (SART) as metrics. SART is the average execution time across instances that all three algorithms can resolve. Researchers present results and discussions from five aspects: acceleration of GPUs relative to CPUs, acceleration of TCFs, memory optimization, comparison with SOTA algorithms, and component acceleration.
Section 8 is the conclusion. In this paper, researchers investigated the use of GPU computing power to solve the MAPF problem. They presented GACB, a parallel two-level algorithm of GPU-acceleration, designed to significantly reduce computational time to solve multi-agent pathfinding problems. The researchers proposed a task coordination framework to facilitate collaboration between high- and low-level solvers simply with lightweight synchronized operations. The researchers also introduced GATSA as a low-level solver for GACBS, allowing efficient parallel processing of SAPF problems under constraints.
The researchers conducted experiments to assess the computational efficiency of the GACBS algorithm on three graphs of various scales. Experimental results demonstrate that GACB significantly accelerates competition-based searches on CPUs, achieving speed-up ratios of over 46 in specific scenarios. Additionally, GACBS surpasses the SOTA reduction-based approach of MAPF problems, achieving speed-up ratios of up to 8.
Gacbs demonstrates CBS's first successful GPU parallelization, but has yet to integrate new techniques that primarily accelerate existing computations and reduce search complexity. Future work will combine CPU-based CBS optimization with GPU parallelization to reduce search complexity and address memory limitations. Additionally, GACBS uses fixed-size arrays instead of dynamic arrays to prioritize computational efficiency. However, GACBS requires array size adjustments to meet the specific needs of your actual application. Future efforts aim to implement dynamic memory allocation mechanisms to increase the adaptability of GACBs.
See the article:
Multi-Agent Emmodied Intelligence's GPU Acrest-based search
http://doi.org/10.1007/S11633-025-1568-y
/Public release. This material of the Organization of Origin/Author is a point-in-time nature and may be edited for clarity, style and length. Mirage.news does not take any institutional position or aspect, and all views, positions and conclusions expressed here are the views of the authors alone.
