Formalizing multimodal knowldge in Traffic Sequence Charts for improving performance, safety, and trustworthiness of AI driving functions

Kurzfassung

Artificial Intelligence (AI) plays an important role in managing complexity of automated driving. Nonetheless, training and ensuring safety of an AI is a challenging task. Especially the safe generalization from known to unknown situations remains an unsolved problem. Infusing knowledge into AI driving functions seems to be a promising approach to address generalization, development costs and training efficiency. A lot of prior knowledge exists for the automotive domain, but its representation differs in quality, type, and format. For instance, mathematical and physical knowledge is often already formalized and can be directly used for AI-Infusion without much effort. Unlike expert knowledge, e.g., traffic rules or court decisions, which are usually not given in a formalized way for integration. Especially in the shift of responsibility for essential driving decisions from humans to AI driving functions, it will be necessary to involve interdisciplinary experts and formalize their knowledge in order to infuse social and societal norms, as well as psychological or physiological aspects. In this presentation, we will present the capabilities of the visual yet formal specification language called Traffic Sequence Charts (TSCs) for formalizing multimodal knowledge, in particular knowledge about traffic maneuvers. We developed an approach using this formalized knowledge to train reinforcement learning (RL) agents, aiming to transform descriptive knowledge on traffic maneuvers in TSCs into performative knowledge in AI traffic agents. We were able to train an agent to control a vehicle through a pass-by maneuver and apply it successfully to an untrained overtaking scenario. In addition, formalized domain knowledge also provides a basis for validating and verifying an AI driving function. To this end, we present an online monitoring for checking the conformance of system behavior against multimodal knowledge formalized in TSCs. With our work, we show that the specification of abstract and multimodal knowledge using TSCs leads to concrete solutions able to support the improvement of AI performance, safety and trustworthiness.