Autonomous Driving via Imitation Learning in a Small-Scale Automotive Platform
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Systems, control and mechatronics (MPSYS), MSc
Publicerad
2024
Författare
Wellander, Johan
Petersén, Arvid
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Abstract
In recent years, the advancement of autonomous driving (AD) technology has garnered significant interest. Traditionally, AD systems have relied on multiple submodules, each handling specific tasks such as perception, path planning, and vehicle control. However, an emerging alternative is the implementation of end-to-end systems, which directly process sensor input to predict vehicle control. While both reinforcement learning (RL) and imitation learning (IL) are utilized
in end-to-end AD systems, RL often finds its strength in simulated environments, where agents learn through exploration and failure. In contrast, IL, learning from an expert model or human, proves more suitable for real-world applications, requiring substantially less data. This thesis presents an implementation of IL for achieving autonomous driving on a go-kart platform. Leveraging both behavioral cloning (BC) and Human Gated Dataset Aggregation (HG-DAgger), we compare the impact of using an interactive IL algorithm HG-DAgger compared to BC. Additionally, our research explores the use of different inputs, including color camera, stereo depth camera, IMU, and the position of ORB features. We also detail the development of a comprehensive software pipeline encompassing data collection, data formatting, model training, and go-kart control. For evaluation, the go-kart was driven around a track for three laps using the trained
BC and HG-DAgger models, and assessed based on number of interventions required per lap, distance without accident, lap time, lap time deviation. The results from the evaluation indicate an improvement in performance from using HG-DAgger over BC as well as an improvement from using a stereo depth camera or the position of ORB features as supplementary inputs to the color camera and IMU.
Beskrivning
Ämne/nyckelord
Keywords: Imitation Learning, Behavioral Cloning, Human Gated Dataset Aggregation, Autonomous Driving, ROS2.