Perception models based on machine learning and deep neural networks are susceptible to misclassifications from subtle perturbations to their inputs. This is concerning because on-road autonomous vehicles (AVs) encounter a variety of perturbations that arise from uncertainty due to the unpredictable behaviors of other actors. Off-road AVs face even greater uncertainty due to the large variation in natural, unstructured environments. Numerous AV simulation tools and algorithms have been created to efficiently explore (and even improve) the robustness of machine learning algorithms for on-road AVs, however, only a handful of tools have demonstrated usefulness for the off-road domain.
This paper presents a modeling and simulation approach for generating adversarial scenes for off-road AVs using reinforcement learning. By “adversarial” we mean that the scene is (ideally) maximally problematic to navigate by the vehicle’s autonomy system while constrained to be realistic. Our work consists of three components: a high-fidelity simulation platform, an adversarial scene generator, and an autonomy system under test. The simulation platform is designed using Unreal Engine 4 and uses a custom plugin to enable users to automatically create basic off-road scenes (e.g., flat ground plane) and run various navigation scenarios. The adversarial scene generator (ASG) uses a distributed Twin Delayed Deep Deterministic Policy Gradient algorithm with prioritized experience replay and a novel action saturation penalty to create test scenarios. The example autonomy system under test has a perception system with a U-Net architecture to predict traversable regions from camera images and uses an A* path planner to avoid the non-traversable regions.
We present results that demonstrate that the ASG architecture can generate pathological scenes on a flat ground plane with up to 32 obstacles. We present studies that highlight various features of the generated scenarios and their implications, and finally we conclude with limitations and future work.
Keywords
AUTONOMY,MACHINE LEARNING,MODELING,SIMULATIONS,SYNTHETIC ENVIRONMENT
Additional Keywords
Off-Road, Reinforcement Learning