Closing the gap between synthetic and real-world domains in autonomous simulation
Xiangyu Bai · 2023
Assessing the mobility performance of autonomous ground vehicles is essential since these vehicles are designed to operate in a wide range of situations and environments. However, relying solely on physical testing is impractical. Therefore, simulations are necessary for their development and evaluation. These simulations must accurately represent reality to enable the vehicles to make decisions as they would in the real world. Unfortunately, existing open-source simulators exhibit a gap between synthetic and real-world domains, which can reduce mobility performance and platform reliability.To address this challenge, our thesis introduces the Scoping Autonomous Vehicle Simulation (SAVeS) platform, which enables benchmarking of simulated environments for autonomous ground vehicle testing. Using SAVeS, we quantify the domain gap between synthetic and real-world domains. Furthermore, we created a more realistic synthetic driving dataset with complex scenes and high-quality textures generated from Grand Theft Auto V's game engine using our Temporal-Controlled Frame Swap (TeFS) method. This dataset provides pixel-accurate dense depth ground truth and stereo vision, both of which are crucial for autonomy research. Finally, we present our efforts to address the domain gap using domain adaptation technologies, including style transfer and depth mask, with SAVeS+. Our results demonstrate that SAVeS+ is effective in closing the gap between synthetic and real-world domains, thereby improving the reliability and performance of autonomous ground vehicles in simulated environments. Overall, the work presented in this thesis is an important step towards realizing autonomy simulation. By developing the SAVeS platform, creating a more realistic synthetic driving dataset, and introducing domain adaptation technologies, we have contributed to advancing the field of autonomous ground vehicle testing and simulation.--Author's abstract