Target Search Based on Scene Priors

Lu Shengyang, Liang Lanjun, Zhao Huailin, Zhou Fangbo, yao Feng

doi:10.5954/icarob.2022.os33-2

Target Search Based on Scene Priors

Lu Shengyang

School of Electrical and Electronic Engineering, Shanghai Institute of Technology
Liang Lanjun

School of Electrical and Electronic Engineering, Shanghai Institute of Technology
Zhao Huailin

School of Electrical and Electronic Engineering, Shanghai Institute of Technology
Zhou Fangbo

School of Electrical and Electronic Engineering, Shanghai Institute of Technology
yao Feng

School of Electrical and Electronic Engineering, Shanghai Institute of Technology

説明

Aiming at the problems of reinforcement learning algorithm in target search tasks, such as low accuracy and low fault tolerance, this article mainly introduces a method of reinforcement learning target search based on scene prior in simulation environment. This method mainly uses graph convolutional neural network to extract the current object relationship as the input of prior knowledge. Secondly, it uses the actor-critic algorithm to take the agent's vision, position and prior knowledge as input to decide the agent's next navigation. Finally, use path planning to navigate to the target point to find the target. Through experiments conducted in Habitat and compared with the previous algorithm, the experiment shows that this method is better than the previous algorithm in target search accuracy and navigation efficiency.