Vertrauenswürdige autonome Systeme

For a climate-neutral and attractive transport mix, the operation of rail transport with the highest levels of automation is an essential component. According to the state of the art, this goal cannot be solved by classical automation technologies alone. On the other hand, there is remarkable progress in the development of technologies in the field of highly automated driving (on road & rail) based on the power of Artificial Intelligence (AI). A major unresolved challenge here is the linkage of AI procedures with the requirements and approval processes in the rail environment.
The use of artificial intelligence (AI) processes – such as machine learning (ML) – is currently a very promising approach to realize automated driving. In a safety-critical environment such as driverless rail transport, robust and safe AI methods are needed to be able to detect obstacles on the track with the necessary reliability.

The safe.trAIn project aims to lay the foundations for the safe use of AI for driverless rail vehicles. Thus, it addresses the greatest technological challenge for the introduction of driverless transport.

Based on the requirements for safety verification, testing methods and tools for AI-based methods are researched. A safety architecture is concretized using the example of the driverless regional train. For this use case, a GoA4 (unattended train operation) system will conceptually be developed and validated in a virtual test field.

The present project aims at the market segment of regional trains. These operate in a more open environment, in which they have to detect obstacles (such as people in the track or trees lying on the track, landslides, etc.) reliably. Compared to road transport, however, there are some factors that make the technical complexity of fully automated rail transport seem lower. Such factors are:

1. fewer interactions with other road users take place in rail operations, which occur especially in areas (e.g., stations or grade crossings) where safety can be reasonably supported by the infrastructure.
2. a collision with other road users on the track is prevented with sufficient certainty by the central control of traffic and automatic cooperation with trackside systems is established.
3. the driving function only has to be provided on a few known routes. This is why the environment can largely be assumed to be known and areas critical to the safety of driverless driving can be additionally secured if necessary.
4. unauthorized entry into the driving path is prohibited. It may even constitute a criminal offense, which greatly simplifies any traffic situations that may arise.

The underlying assumption of the project is therefore that various requirements are technically easier to solve in rail-bound traffic compared to road traffic. Thus, a timely market introduction appears more realistic. On the other hand, there are considerable differences both from a regulatory point of view and from a technical point of view (e.g., the need to use sensor technology with a longer range to observe the track, availability of suitable training data). So, the solutions, developed in the automotive sector, cannot be adopted directly. However, the ongoing research projects in the automotive industry using AI for safe environment recognition represent important preliminary work that will be taken up in this project and further developed for use in the rail sector.