Zero-Shot Policy Transfer in Autonomous Racing: Reinforcement Learning vs Imitation Learning

There are few technologies that hold as much promise in achieving safe, accessible, and convenient transportation as autonomous vehicles. However, as recent years have demonstrated, safety and reliability remain the most obstinate challenges, especially in complex domains. Autonomous racing has demonstrated unique benefits in that researchers can conduct research in controlled environments, allowing for experimentation with approaches that are too risky to evaluate on public roads. In this work, we compare two leading methods for training neural network controllers, Reinforcement Learning and Imitation Learning, for the autonomous racing task. We compare their viability by analyzing their performance and safety when deployed in novel scenarios outside their training via zero-shot policy transfer. Our evaluation is made up of a large number of experiments in simulation and on our real-world hardware platform that analyze whether these algorithms remain effective when transferred to the real-world. Our results show reinforcement learning outperforms imitation learning in most scenarios. However, the increased performance comes at the cost of reduced safety. Thus, both methods are effective under different criteria.