Dataset of Distributed Neural Policy Gradient Algorithm for Robot Path Planning: This dataset was established in November 2024 and generated using a Python simulation platform, primarily targeting performance analysis and verification of distributed neural policy gradient algorithms in robot path planning tasks. The dataset contains numerical simulation results of centralized and distributed algorithms across two types of path planning scenarios, used to verify the convergence performance, cooperative decision-making capability, and computational efficiency of the proposed algorithm in complex path planning environments. The dataset is constructed based on a multi-robot path planning environment, generating dynamic evolution data of different algorithms during the training process through numerical simulation of robot state transitions, action selections, and environmental interactions. The objective function performance metric is composed of the reward for reaching the target point, the collision penalty incurred between robots or between robots and obstacles, and the time penalty during the path planning process, used to evaluate the comprehensive performance of the algorithm in terms of task completion quality and safety. Meanwhile, the dataset records the norm variation of the objective function gradient with respect to policy parameters, used to analyze the optimization stability and convergence characteristics of the neural policy gradient algorithm. In addition, the runtime of different algorithms during training and decision-making processes is also recorded, used to compare the differences between centralized and distributed methods in terms of computational complexity and real-time performance. All dataset files are in NPY format, comprising 12 files in total, organized by algorithm type and scenario index, with the naming convention "CTP_PathStructure_AlgorithmTypeDataType.npy", where path structure is identified by numerical index representing different planning scenarios (11 and 321), algorithm type is divided into centralized (cen) and distributed (dis), and data type includes objective function value (obj), policy gradient norm (gra), and runtime (runningtime), specifically including CTP_11_cenobj.npy, CTP_11_cengra.npy, CTP_11_cenrunningtime.npy, CTP_11_disobj.npy, CTP_11_disgra.npy, CTP_11_disrunningtime.npy, CTP_321_cenobj.npy, CTP_321_cengra.npy, CTP_321_cenrunningtime.npy, CTP_321_disobj.npy, CTP_321_disgra.npy, and CTP_321_disrunningtime.npy. The dataset contains 3 categories of parameters: objective function value (dimensionless, reflecting the comprehensive score of task rewards and penalties), policy gradient norm (dimensionless, reflecting the magnitude of policy parameter updates and used to measure algorithmic convergence stability), and runtime (in seconds, reflecting the computational cost of the algorithm during training and decision-making), with data classified along two dimensions of algorithm type (centralized/distributed) and path scenario (Scenario 11/Scenario 321). In the course of the research, this dataset is used to support the performance verification and effectiveness evaluation of the distributed neural policy gradient algorithm, validate the effectiveness of the proposed distributed policy update mechanism in multi-robot cooperative path planning, and analyze the performance of the algorithm under different communication and cooperative structures in terms of path optimization, collision avoidance, and training efficiency, thereby meeting the simulation verification requirements for robot swarm path planning and cooperative control in complex dynamic environments.
This dataset consists of simulation-generated data and was not obtained from literature, mirroring, purchase, or external download. Details are as follows: The data was generated by the research group based on a self-developed multi-robot path planning simulation environment built on the Python simulation platform (open-source code available at https://github.com/Pengcheng-Dai/DACA), through numerical simulation training and testing of centralized algorithms and distributed neural policy gradient algorithms. Equipment: A networked computer running a Python simulation environment (NumPy for numerical computation, policy updates, and gradient calculations; Matplotlib for plotting and visualization of results). Methods and Process: Under two different multi-robot path planning scenarios, numerical simulations were conducted on robot state transitions, action selection, and environment interaction processes. Centralized and distributed neural policy gradient algorithms were run respectively, with the following data recorded during training: objective function performance metrics (reward for reaching the target, collision penalties, and time penalties), the norm of the gradient of the objective function with respect to policy parameters, and algorithm runtime, which were automatically generated as the corresponding simulation dataset. Data Collection Time and Location: November 2024, completed on the laboratory computing platform.
Data Processing Methods: This dataset consists of numerical data generated through reinforcement learning algorithm model simulations. The specific processing methods are as follows: Simulation Environment Construction: A multi-robot path planning simulation environment was built on the Python platform, modeling the robots' state space, action space, reward function, and obstacle distribution. Algorithm Models: Centralized policy gradient algorithms and distributed neural policy gradient algorithms were used to train robot path planning policies. The policy network is represented as a neural network, with policy parameters iteratively updated via the policy gradient method. Building on this, the distributed algorithm introduces a mechanism for local information exchange and cooperative updates among multiple agents. The detailed simulation environment structure can be found in the research group's open-source project at https://github.com/Pengcheng-Dai/DACA. Mathematical Operations and Metric Calculations: The objective function performance metric is calculated as a weighted sum of three components: the reward for reaching the target point, the collision penalty (for collisions between robots or with obstacles), and the time penalty, reflecting the overall performance of each episode; The policy gradient norm data is obtained by computing the gradient of the objective function with respect to the policy parameters and calculating the Euclidean norm (L2 norm) of this gradient vector, used to measure the convergence trend of the optimization process; The runtime data is obtained by timing the training and decision-making computation process for each iteration (or each episode) of the algorithms. Data Recording and Processing: NumPy was used to perform numerical calculations and statistical aggregation of the various metrics generated during the simulation, which were then stored as numerical arrays categorized by algorithm type (centralized/distributed) and scenario type. Finally, Matplotlib was used to plot convergence curves, performance comparison curves, and runtime statistical charts.
All numerical indicators in this dataset (such as the performance value of the objective function, the policy gradient norm, the running time, etc.) are stored in the form of double-precision floating-point numbers (float64), directly output by the Python and NumPy numerical computing environments, retaining many significant digits after the decimal point, without any manual truncation or rounding processing. It can truly reflect the numerical changes during the simulation process.
This work is licensed under
CC BY 4.0 (Creative Commons Attribution 4.0 International License).
