Visual Language Navigation Based On Knowledge Integration Of Indoor Scene Graph

Visual language navigation involves a comprehensive understanding of visual images and natural languages,as well as path optimization,which is challenging.As an important step for intelligent robots to enter human life,visual language navigation is becoming a hot research topic in the field of artificial intelligence in recent years.By learning to navigate independently,the robot can get rid of its dependence on environmental maps,deploy faster to a variety of tasks and facilitate task expansion.Traditional algorithms use the method of designing navigation controller after modeling,which makes it difficult to deal with.With the development of deep reinforcement learning,most of the current research work uses simulated learning or reinforcement learning to solve visual navigation problems.Visual language navigation is essentially a problem of target search and positioning.Robots’ awareness of the search environment is an important factor to improve the search efficiency of algorithms.In order to automatically build a priori knowledge suitable for the target environment,this paper uses scene graph generation combined with graph convolution network learning to build a knowledge graph,which integrates the knowledge information into the deep reinforcement learning framework.In order to further improve the model’s adaptability to the strategy and prior knowledge in reinforcement learning during the testing phase,this paper adds a Model-agnostic meta-learning process,put trainable self-supervised interaction loss into the navigation reinforcement learning framework,and improves the generalization ability of the model.Therefore,we can acquire corresponding prior knowledge in different scenarios to assist our visual navigation tasks.This paper conducts experiments in the mainstream AI2-THOR simulation environment.Compared with the standard reinforcement learning method and only the manually constructed knowledge graph method,this model has superior navigation efficiency and good knowledge automation extraction and construction ability,which verifies the validity of adaptive prior knowledge.At the same time,to verify the model’s performance in different scenarios,we test the model in another Matterport3 D,which is also widely used.Finally,it is proved that our model can also effectively improve navigation efficiency in different scenarios.