场景感知的分布式多智能体目标搜索方法

Scene-aware decentralized Monte Carlo Tree Search of target discovery

在视觉语义导航任务中,智能体通过视觉信息,寻找并导航到给定对象类别的目标处。然而,大部分现有的研究都是使用基于学习的框架来完成任务,这些研究在现实世界中应用的训练成本非常高,可移植性很低,并且它们只适用于单智能体,效率低下、容错能力差。为解决上述问题,本文提出一种基于场景感知的分布式多目标优化蒙特卡洛树搜索模型,该模型中多智能体实时在线规划并且不需要预先训练,利用场景感知先验知识结合观测信息实时对环境进行估计,并且利用改进的蒙特卡洛树搜索进行路径规划以此搜索目标。在Mtterport3D数据集中进行的实验表明,该模型在效率方面比单智能体有着显著的提高。

In visual semantic navigation,agents find and navigate to the target of a given object category through visual information.However,the majority of existing studies complete the task using a learning-based framework.These studies have a high training cost in the real world,low portability,and are only suitable for single-agent systems with low efficiency and poor fault tolerance.To address the above issues,this paper suggests a decentralized multi-objective optimization Monte Carlo Tree Search model based on scene awareness.In this model,multi-agent plans online in real time and does not need to train in advance.The improved Monte Carlo Tree Search is used for path planning to search targets,and the environment is estimated in real-time by using scene awareness prior knowledge combined with observation information.Experiments in the Matterport3D dataset demonstrate that the effectiveness of the model is significantly higher than that of a single agent.