Constantly changing domains represent a significant challenge for the application of artificial intelligence (AI) in the automotive context and especially for autonomous driving. AI solutions for autonomous driving must be responsive to an ever evolving market and must be scalable to meet changing requirements; a capability known as Autonomy at Scale. Typical examples of domain change are different use cases, sensors, ECUs, and changes in time and location. To improve development efficiency for changing application domains, the question must be asked:

Is it possible to transfer knowledge gained from previously learned domains to the requirements of new target domains and to concentrate exclusively on learning the delta, i.e. on the essential domain differences?

The aim of the AI-Delta Learning project - a project of the VDA lead initiative "Autonomous and networked driving" - is to develop methods and tools for the efficient extension and adaptation of existing AI modules of autonomous vehicles to the requirements of new areas and increasingly complex scenarios. The methods developed will make it possible to efficiently transfer existing knowledge from one domain to applications in new domains. Only additional requirements, the deltas, will then be re-learned with minimal development effort.

In order to master these tasks, a consortium of leading OEMs, automotive suppliers, technology providers as well as many universities and research institutions has joined forces and started work with the kick-off event on 22 and 23 January 2020.

In the project context, OFFIS will deal with questions concerning the generation and use of synthetic data for the learning process as well as the robustness evaluation of AI models. Thereby an efficient bridging of deltas shall be enabled. Furthermore, we will also consider problems on the hardware-software level, since the systems embedded in a vehicle are subject to quite strict requirements, which have a transitive effect on the possible complexity of the AI models. OFFIS will develop methods to evaluate and optimize models regarding different factors (real-time capability, memory requirements, …).