Generating and Explaining Corner Cases Using Learnt Probabilistic Lane Graphs
This addresses the problem of insufficient training data for safety-critical scenarios in autonomous vehicles, though it is incremental as it builds on existing reinforcement learning and graph-based methods.
The paper tackles the challenge of validating autonomous vehicle safety by generating novel, realistic corner case scenarios using Probabilistic Lane Graphs (PLGs) learned from traffic data, enabling explainable safety-critical situations for simulation-based testing.
Validating the safety of Autonomous Vehicles (AVs) operating in open-ended, dynamic environments is challenging as vehicles will eventually encounter safety-critical situations for which there is not representative training data. By increasing the coverage of different road and traffic conditions and by including corner cases in simulation-based scenario testing, the safety of AVs can be improved. However, the creation of corner case scenarios including multiple agents is non-trivial. Our approach allows engineers to generate novel, realistic corner cases based on historic traffic data and to explain why situations were safety-critical. In this paper, we introduce Probabilistic Lane Graphs (PLGs) to describe a finite set of lane positions and directions in which vehicles might travel. The structure of PLGs is learnt directly from spatio-temporal traffic data. The graph model represents the actions of the drivers in response to a given state in the form of a probabilistic policy. We use reinforcement learning techniques to modify this policy and to generate realistic and explainable corner case scenarios which can be used for assessing the safety of AVs.