Object-oriented indoor navigation for delivery robot

Abstract

The navigation of object-oriented delivery robot is widely used in daily life. This thesis focuses on the popular VLN tasks in recent years to solve the problem of indoor delivery navigation in unseen environments. In the research of VLN, the cross model interaction between vision and language has made significant progress in the past two years with the rapid development of CV and NLP. The emergence of BERT models also help in training and construct ing navigation frameworks. Although the BERT model has good performance in VLN, the mismatch between instructions and visual information at the input leads to navigation errors for robots in similar scenes. This thesis introduces a cross model interaction transformer to solve the mismatch between instruction and visual information to optimize the input of the BERT model and improve the navigation success rate of the delivery robot.