Continuously changing domains pose a significant challenge to the application of Artificial Intelligence (AI) in the automotive context in general and its application to autonomous driving in particular. AI solutions for autonomous driving have to respond to a constantly evolving market and must be scalable to meet changing demands; a capacity referred to as Autonomy at Scale. Typical examples of domain change include varying use cases, sensors, ECUs along with development states, as well as changes in time and location. To improve the development efficiency for changing application domains the question must be raised:

Is it possible to transfer gained knowledge from previously learned domains to the demands of new target domains and to solely focus on the learning of the delta; that is, the main domain differences?

Illustration 1: Autonomy at Scale

The goal of the KI Delta Learning project is to develop methods and tools for the efficient extension and adaptation of existing AI modules of autonomous vehicles to meet the demands of new domains and increasingly complex scenarios. The developed methods will allow existing knowledge from one domain to be efficiently transferred to applications in new domains. Only additional requirements, the deltas, are then to be re-learned with minimal development effort.

Illustration 2: Domain Changes in Delta Learning


Within the project KI Delta Learning 19 partners from industry and research are developing benchmarks for different cases of delta-learning and will provide new algorithmic solutions within the following three main categories:

Illustration 3: The Principle of Delta Learning

Transfer Learning: Deltas between previously trained domains and new domains.

Didactics: Deltas in learning processes and training strategies.

Automotive Suitability: Deltas between AI solutions for automotive requirements and existing AI applications from other fields.

Existing approaches that will be further developed within the project’s context of delta-learning for autonomous driving include:

Continues Learning, Active Learning, Novelty Detection, Anomaly Detection, Semi-Supervised Learning, Unsupervised Learning, One-Shot Learning, Augmentation, Catastrophic Forgetting, Knowledge Distillation and Graph Networks.

Project Innovations:

  • Learning strategies and transfer learning methods
  • Methods for generating, annotating and evaluating training data used in delta learning
  • Technologies to manage and process acquired data
  • Methods for automatic training process optimization

The Project at Glance

Total budget: EUR 27.44 million

Duration: 01.01.2020 – 31.12.2022

Consortium leader: Mercedes-Benz AG


Contact us

Coordinator and Project Lead:

Mohsen Sefati
Mercedes-Benz AG
Urban Automated Driving
Hessbruehlstrasse 21
70565 Stuttgart-Vaihingen

Project Manager:

Dr. Stefan Dietzel
+49 30 3670235-136

European Center for Information and Communication Technologies – EICT GmbH
EUREF-Campus Haus 13
Torgauer Straße 12-15
10829 Berlin

KI Delta Learning is a part of the AI project family of the VDA Leitinitiative Autonomous and Connected Driving