Behavior-Tree Based Scenario Specification and Test Case Generation for Autonomous Driving Simulation

摘要

Autonomous driving systems are notoriously difficult to be made safe, largely attributed to their complex and dynamic environments, also known as scenarios. Simulation is an effective way to find uncovered safety bugs and compare autonomous driving algorithms through a large number of concrete scenarios (or test cases). The behavior tree is an established formalism for describing and controlling the behaviors of actors in game AI as well as robotics. In this paper, we propose a new approach based on the Behavior Tree for scenario specification and test case generation. We propose a scenario description language, called BTScenario, which supports both the high-level description of the spatial layout of map objects (e.g. roads and junctions) and the Behavior Tree-based specification of temporal behaviors of actors (e.g. cars and pedestrians). Furthermore, we implement the BTScenario program with the off-the-shelf opensource Behavior Tree library, PyTrees, to automatically generate test cases for simulation platforms. We test driving areas on maps to find bugs and compare different control algorithms by a large number of test cases generated automatically.