Introduction

Reliable perception plays a key role for mobile robots to navigate robustly in unstructured environments such as wild, battlefield, and other planet’s surface. Recently, machine learning methods have shown great usefulness and brilliant foreground in the field of robotic perception. However, most current machine learning based perception methods require a large number of labeled data for training. Unfortunately, labeling multi-sensor data is time consuming and costly. So, this project plans to address the perception problem by introducing active learning techniques to reduce the need for large quantities of labeled data. The main contents of this project include the following four parts: data preprocessing, automatic selection of unlabeled data for labeling, supervised learning based on labeled data, and evaluation of proposed methods. The project is expected not only to solve some basic science problems in active learning but also to develop an effective obstacle detector and terrain classifier.

[ GO TOP ]