A Survey: Vision-based Mapless Navigation for Ground Robots with Deep Reinforcement Learning

PhD Qualifying Examination


Title: "A Survey: Vision-based Mapless Navigation for Ground Robots with 
Deep Reinforcement Learning"

by

Mr. Xiaodong MEI


Abstract:

Autonomous mapless navigation is one of the most significant problems for 
ground robots. The task is to guide the robot to reach a certain goal 
location timely without obstacle collisions. Since the mapless navigation 
system does not rely on the provided map and map-construction process, it 
makes decisions as it perceives the environment, which is easy to be 
formulated into deep reinforcement learning (DRL) frameworks. As vision 
becomes more and more common in the ground robot researches, in this 
survey, we mainly focus on the vision-based navigation systems with DRL 
algorithms, especially for the ground robot in indoor and structured 
outdoor environments, such as the urban areas. We first define the mapless 
navigation problem and its challenges. Then we introduce how the problem 
can be formulated into DRL framework. We briefly review the basic concepts 
and algorithms of DRL, especially the model-free ones. Next, we introduce 
the representative works of vision-based mapless navigation with DRL. For 
each method, we analyze the contributions and limitations. And we 
introduce the reality gap problem, which is the main challenge of the 
vision-based navigation system. Based on the summary of the previous 
methodology, we propose some promising directions for further research.


Date:			Thursday, 18 June 2020

Time:                  	10:00am - 12:00noon

Zoom meeting:           https://hkust.zoom.us/j/98955446213

Committee Members:	Dr. Ming Liu (Supervisor)
 			Dr. Qifeng Chen (Chairperson)
 			Prof. Dit-Yan Yeung
 			Prof. Tong Zhang


**** ALL are Welcome ****