PGDrive: Procedural Generation of Driving Environments for Generalization
Quanyi Li,Zhenghao Peng,Qihang Zhang,Chunxiao Liu,Bolei Zhou
2021-01-01
Abstract:A dilemma in prototyping autonomous driving systems is that, due to public safety concerns, it is impossible to directly develop, train, and test the self driving systems in the physical world [13]. To overcome the dilemma, a common practice is to prototype and validate the driving system in simulators. Many realistic driving simulators have been proposed. CARLA [4] and SUMMIT [1] contain driving scenarios with realistic visual appearances. Flow [14], Duckietown [2], and Highway-env [7] either focus on learning high-level decisions or only contain simplistic driving scenarios. However, it’s notorious that neural networks can overfit training data easily [15, 3], thus the driving agent trained on a fixed set of scenarios in the simulators might have difficulty generalizing to new scenarios. Furthermore, the existing simulators lack sufficiently diverse designs of the maps and scenarios to evaluate such issues. For example, CARLA consists of only about ten manually designed towns, which is hard to modify, while Highway-env contains six typical traffic scenarios. Training deep neural networks on such simulators might lead to overfitting and poor generalization of learned driving policy. To evaluate and improve the generalization of end-to-end driving, we introduce PGDrive, an open-ended and highly configurable driving simulator. The key feature of PGDrive is Procedural Generation (PG) [12, 10]. Unlimited number of diverse driving scenes can be generated with the help of our abstraction on driving scene as well as the proposed PG algorithm. PGDrive is highly configurable, allowing users to customize settings such as the traffic flow and the vehicles’ dynamics models. Besides, fine-grained RGB camera, depth camera, Lidar data generation, and realistic 3D kinematics simulation are also supported. To support fast prototyping of driving system, PGDrive is designed to be