TESTING AUTONOMOUS VEHICLES USING REINFORCEMENT LEARNING TO GENERATE FAILURE SCENARIOS IN COMPLIANCE WITH STANDARDIZED TESTS

Abstract

This thesis proposes a design for a reinforcement learning framework to test specific autonomous vehicle components according to standardized tests of EuroNCAP. It shows how reinforcement learning algorithms are being used in real-world applications, in different testing domains outside the autonomous vehicle testing, and how to make use of reinforcement learning algorithms for autonomous vehicle testing rather than the popular topic of usage in driving autonomous vehicles. In addition, it presents a complete reinforcement learning formulation for the framework including environment description, reward function design, model training, and model testing procedures. Moreover, the proposed framework was able to generate automatic failure scenarios that were applied on autonomous vehicles covering two EuroNCAP scenarios; approaching a stationary car and approaching a slower car. The proposed framework controls parameters such as velocity, position and time, and generates more accurate failure scenarios to happen in real-life situations. Our failure scenarios are generated using q-learning and deep reinforcement learning algorithms causing real accidents for the designed scenarios. Hence, our reinforcement learning framework proves its validity to generate failure scenarios for autonomous vehicle components improving the safety of autonomous vehicle components and reducing both the costs and time required for testing autonomous vehicle components.