KI Delta Learning

Scalable AI for Automated Driving

Final project results

Please go to Publications and navigate to Results for project results.

Efficient training methods for scalable, highly automated driving

Highly and fully automated vehicles are expected to become an integral part of traffic. However, they face a large variety of complex traffic situations in a continuously evolving environment. This leads to great challenges for the development of autonomous vehicles ready for serial production.

Apart from dealing with changes in software and hardware, automated vehicles must learn to cope with constantly changing situations, and therefore with different so-called domain shifts.

The KI Delta Learning project investigates methods of effective machine learning that enable a more efficient and unlimited use of artificial intelligence (AI) and, hence, the implementation of automated driving in the "open world". Thanks to the methods developed for this purpose, automated vehicles can be faster deployed in new markets and react more agilely to new requirements.

Training AI and defining deltas

In the KI Delta Learning project, leading partners from science and industry deepen the expertise around the successful use of AI. Only efficient training of AI modules on diverse domains ensures that the approach can be applied globally and that autonomy at scale turns into reality.

Existing domains are trained for additional requirements. The knowledge is optimally expanded.

Here, the focus is on transferring the knowledge already derived in known domains to new target domains and on efficiently learning the additional requirements of changed application areas, the deltas:

Changes and further developments in the area of sensors
Diverse traffic areas – from simple country roads to highly complex inner-city traffic routing
National peculiarities in traffic routing
Temporary changes, such as daytimes and seasons
Long-term changes in traffic zones with new concepts of mobility and road users
Continuous development of AI methods, such as better training strategies and more efficient neural networks