Data-driven traversability estimation for mobile robot navigation

Abstract

Mobile robots have a promising application prospect as they can assist or replace humans to perform laborious, repetitive or dangerous tasks in various scenarios. There has been a large number of studies for mobile robot navigation since 1980s, while terrain traversability estimation is an important topic in this field — estimating if an area is traversable and how long will it take to drive through is necessary for navigating the robot and planning paths. However, most existing terrain traversability estimation methods are designed based on simple fixed rules and manually tuned parameters, suffering low accuracy

and poor generalization due to their simplicity of structure and biases to the environ- ment where they are tuned. To address this problem, we proposed a set of data-driven

traversability estimation methods based on Convolutional Neural Networks (CNN), which are trained and tested them in different simulation environments. There are 3 main goals for our methods: 1. High accuracy. Accuracy of the result is the core of an traversability estimation method.

2. Low computational cost. Since most mobile robots are equiped with very lim- ited computing power and energy, a practical traversability estimation method

should be able to work with a low computational cost. 3. Good generalization. A good traversability estimation method should generalize to different type of environments or provide a function to automatically fit to a new environment instead of manual tuning. In this thesis, we first reviewed some representative conventional terrain

traversability estimation methods and introduced several related fields including mo- bile robot path planning, localization and map building. Then we proposed our

CNN-based methods, demonstrated how to build the simulation framework and col- lect terrain samples with driving data. Finally we compared the performance of our

work with benchmark methods in both classification and regression traversability es- timation tasks on the collected datasets and proved the improvements made by our

methods.