Generation of naturalistic and adversarial sailing environment (NASE) for intelligent test of autonomous ships

Abstract

—Intelligence testing and verification of autonomous sailing capability is crucial to the development and deployment of autonomous ships. The prevailing approach is to establish a real test site on the sea or create a life-like simulator to test the autonomous sailing functions. However, due to the high dimensionality of the naturalistic sailing environment and the rareness of encounter events, hundreds of millions of miles and long test cycles would be required to demonstrate the collision avoidance performance of autonomous ships, which results in huge consumption. A novel approach is presented to generate the naturalistic and adversarial sailing environment by employing the sparse but adversarial adjustments into the naturalistic sailing environment. Our method can be used to enhance the existing sailing environment simulations to accelerate the autonomous ship test process. Firstly, the naturalistic probabilistic distributions of ship sailing behaviors are analyzed. Then, the naturalistic sailing environment (NSE) is generated to achieve real-world renderings. Moreover, the characteristics of ship maneuvering behavior at different encountered situations are identified. Furthermore, the naturalistic and adversarial sailing environment (NASE) is established by sampling maneuvers of the background ships (BS) from the importance functions at critical moments. It ensures the authenticity of the scenarios and increases the generation probability of encounter scenarios. The effectiveness of the proposed naturalistic and adversarial sailing environment model in the simulation of narrow waterway environment is proved. Compared with the naturalistic sailing environment, the proposed model can accelerate the generation speed of the critical scenarios and improve the generation efficiency of encounter scenarios.