Evolutionary Single Shot Tokenized Gan for Simulation Environment Generation and Fault Detection

Advanced robotic technologies such as self-driving cars, autonomous robots, and unmanned aerial vehicles require rigorous testing to expose flaws in their control software. Developing suitable testing environments is a challenging task due to the difficulty of designing adequate testing environments that uncover faults in the robotics control software. In this paper, we create a custom Simulation Evolutionary Single-Shot Generative Adversarial Network titled SESS-GAN. It makes use of single shot tokenized GANs to generate a preliminary simulation environment model, which is then fed into an evolutionary algorithm for finetuning. This process increases the risk score of the simulations; meaning these environments will cause the autonomous robot control software to fail/near fail, allowing testers to evaluate its performance accordingly. Empirical results highlight the potential of our solution with multiple use case environments.