Cooperative Navigation Benchmark Environment

该数据集涉及多智能体协作导航任务，在二维网格世界中，智能体们需要各自导航以达到设定的目标。奖励函数定义为智能体与目标之间的欧几里得距离的负值，同时对于碰撞行为设有惩罚。智能体的策略由一个包含三层的密集神经网络来表示。该任务的规模涉及五个智能体，属于多智能体强化学习任务范畴。

This dataset focuses on multi-agent collaborative navigation tasks. In a 2D grid world, each agent is required to navigate independently to reach its preset target. The reward function is defined as the negative Euclidean distance between an agent and its target, with an additional penalty for collision behaviors. The policy of each agent is represented by a three-layer dense neural network. This task involves five agents and falls under the category of multi-agent reinforcement learning tasks.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于协作导航基准环境的代码库，实现了MDPGT（基于动量的去中心化策略梯度跟踪）算法，支持Lineworld和Particle-world两种环境，并包含DPG、MDPG和MDPGT三种代理训练方法，用于复现相关论文的研究结果。