RTAS: Road Test with Artificial Scenarios

Abstract

The decision making module in self-driving cars requires extensive performance testings and safety evaluations on real roads. While providing a naturalistic driving environment, the road test poses several problems, including uncontrollable risks, rare complex and dangerous scenarios, and higher costs. Testing in a simulator could provide unrestricted scenarios, but it lacks real-time dynamic feedback from either vehicle or road. To bridge the gap between physical and virtual testing, we propose a novel test and evaluation system, called Road Test with Artificial Scenarios (RTAS), which injects generated virtual scenarios to the physical VUT (Vehicle Under Test) in real time. For virtual scenarios, we propose a deep generative network with the road structure, layout of traffic participants and expected safety critical measurement as inputs. Meanwhile, the VUT is driving in a controlled physical environment (e.g. a test track) and its motion planner is modified by directly taking generated scenarios as inputs. Furthermore, the motion of the VUT is captured by localization devices and passed to the scenario generation as the instant feedback of the VUT. To demonstrate the feasibility of our proposal, we implement a prototype based on a scaled indoor test field, which integrates our scenario generation with Carla simulator and a 1/18 scale vehicle running on a scale indoor test field.