Using Machine Learning to Improve Autonomous Robot Navigation on Vertically Challenging Terrain

Abstract

Autonomous robots can easily navigate flat surfaces like roads and sidewalks; however, when it comes to vertically challenging terrain, such as jagged rocks in off-road conditions, current robots tend to struggle. In order to address this problem, our research utilized machine learning (teaching a robot to make optimal decisions based on expert human demonstration). We built a four-wheel robot (crawler), constructed from a Traxxas chassis, with the following elements added to it: an Arduino Mega 2560, a Microsoft Azure Kinect RGB-D camera, an NVIDIA Jetson ORIN NX onboard computer, and a 6-volt servo. We then used this crawler to collect data by manually driving it over a bed of rocks and recording an elevation map generated by the RGB-Dcamera. In addition to mapping, each time we drove the robot over a bed of rocks, we recorded factors such as speed, odometry, and the state of its differentials (either locked or unlocked). Our lab will generate a neural network with the data collected and use it to train the crawler to traverse similarly difficult terrain autonomously in the future. By teaching it to recognize the best paths and courses of action on such terrain, the possibility of the robots’ ability for autonomous navigation creates more opportunities for the application of robotics in the real world.