人形機器人騎乘電動機車時之視覺里程計

Abstract

noneIn recent years, deep learning has been used to develop autonomous driving on four wheels, with the goal of reaching autonomy. Our laboratory is dedicated to developing intelligent humanoid robots and is willing to take on new research fields - autonomous driving of an unmodified humanoid robot on two wheels. Taiwan has a well-established scooter industry, but few researchers have studied the behavior of autonomous scooter driving. Our ambitious plan is to use a large humanoid robot called Thormang3 to develop an autonomous scooter system and attempt to pass the driving test for the Taiwanese scooter license. To achieve self-balancing in a real environment, the current speed detection and control of the scooter are crucial issues. The main contribution of this paper is a speed controller for a two-wheeled electric scooter, using a large humanoid robot to achieve constant speed driving in a real-world environment. Currently, we are using three main methods to obtain the current speed of the scooter: Yolo dashboard speed detection, ORB SLAM3, and a hybrid method.We will evaluate the accuracy of these methods in an outdoor environment and discuss their advantages and limitations. By using the linearity of the speedometer, we can obtain a rough velocity estimate for the robot using Yolov4 object detection. During the robot's navigation, the rough velocity estimate provides a relatively accurate measure of the real-world scale factor necessary for ORB SLAM3, which helps overcome the inherent disadvantages of monocular cameras and improve real-time velocity extraction.