Scalable and learnable autonomous driving system : design and field evaluation

Abstract

Autonomous driving systems can be applied in a variety of conditions, from complex urban environments and multi-story buildings to high-speed racing settings. However, developing a system specialized for each environment requires substantial resources and effort. To address this challenge, our research proposes an autonomous driving system that is both scalable and learnable. The system architecture is modular, composed of navigation, perception, decision-making, and control. Through independent development and evaluation of each module, we enhanced the overall coherence and efficiency of the system. The navigation module, in particular, is pivotal for the operation of autonomous systems. Our research delves into the intricate challenges of navigation in multi-story buildings, high-speed racing environments, and urban settings where GPS functionality is constrained. Initially, to enhance reliability, we designed a multi-modal system that merges GPS, Camera, and LiDAR sensors. By employing probabilistic models, we can identify and rectify sensor discrepancies or performance degradation, allowing for refined navigation solutions. Furthermore, to deal with GPS-denied environments, we carried out research on robust state estimation by solely utilizing LiDAR sensors with 3-D high-definition map. In addition, we discuss methods to maximize computational efficiency, ensuring operability even in edge computing systems such as racecar and mobile robots. Lastly, leveraging repetitive experimental data, we have constructed an evolvable model and framework that facilitates position estimation through learning, thus crafting a navigation algorithm adept at adapting to evolving surroundings. On another note, for consistent driving across diverse terrains, utilizing a node-link-based road model for vehicles or mobile robots can be highly beneficial. Hence, our research also encompasses methods for automatic road model graph generation, alongside strategies for route planning and optimal path selection, ensuring the system can smoothly navigate around various obstacles encountered during its journey. Conclusively, we applied the proposed system to diverse platforms like mobile robots, urban autonomous vehicles, and racing cars. Each developed module was evaluated through an array of scenarios, spanning from simulations to real-world environments.