Rail
Safe.trAIn: AI as train conductor on regional railways

Autonomous driving is not only set to conquer the roads, but also offers enormous potential for driverless trains in rail transport. Proof of this is the safe.trAIn project, in which Fraunhofer IKS has taken on the central task of methodically assuring the AI functions.

September 4, 2025

I Stock 1208411270 Thomas Stockhausen
mask I Stock 1208411270 Thomas Stockhausen

A key driver for the development of autonomous trains is the looming shortage of qualified train drivers. Studies indicate that in the coming years, there will be significantly fewer staff available than would be necessary for stable and reliable rail operations. Driverless systems therefore offer the opportunity to close this gap, increase the efficiency of rail traffic, and even (re)open previously unserviceable routes in an economically viable manner. Driverless operation (also referred to as Automatic Train Operation, ATO) is thus becoming a key technology for making rail transport future viability and sustainability.

The safe.trAIn project can be considered a significant milestone on the path to fully automated trains in open operating areas such as regional transport. The aim was to research the safe application of artificial intelligence (AI) based functions in driverless rail transport – in particular for obstacle detection in so-called GoA3 to GoA4 systems, i.e., fully automated operation without a driver or HR on board (see info box). The Fraunhofer Institute for Cognitive Systems IKS played a central role in the methodological assurance of the AI functions.

Safetrain Grafik 02 01
Bild

Fig. 1 : Schematic representation of the contributions in safe.trAIn: (1) Safety concept), (2) AI metrics, and (3) Rail ODD

A key outcome of the project was the development of a systematic safety concept for AI-based functions. Researchers at Fraunhofer IKS led the safety case development based on a method called Goal Structuring Notation (GSN), a graphical representation that brings together safety-related evidence in a structured manner. Using a methodology called Landscape of Safety Concerns, AI-specific risks were taken into account based on identified AI weaknesses, suitable metrics for evaluating these risks were developed, and integrated into the safety argumentation.

Focus on AI metrics

Particular attention was paid to the development and evaluation of AI metrics. The researchers were responsible for analyzing existing metrics to determine whether they were suitable for evaluating key characteristics such as trustworthiness, explainability, and safety performance. To this end, the research scientists at Fraunhofer IKS also developed their own metric called PROWL (Prototype-based Out-of-Domain detection Without Labels).

PROWL enables zero-shot detection of objects outside the defined operating conditions (the so-called Operational Design Domain, ODD) – a crucial contribution to the safety of AI systems in the railway sector.

On behalf of the Fraunhofer-Gesellschaft, also the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS was involved in the project, contributing to the project's success with the "Semantic Performance Discrepancy" metric and the ScrutinAI tool. Both approaches enable the systematic identification of weaknesses in deep learning models and support the visual analysis of safety-relevant AI components – an important contribution to trustworthiness and explainability.

Safetrain Grafik 2
Bild

Fig. 2 : Top-level overview of the Rail ODD taxonomy


Another focus was on the aforementioned Operational Design Domain (ODD), i.e., the definition of the conditions under which an AI system can be operated safely. Fraunhofer developed an ODD taxonomy for rail (see Figure 2) as well as an associated procedure and description methodology. This was standardized under the chairmanship of Fraunhofer IKS in the DIN DKE SPEC 99004 standard and now forms the basis for future AI applications in railway.. With the specially developed ODD Workbench (see "Railway AI Systems: The Importance of Operational Design Domain (ODD") a software tool has also been created that enables the structured creation, management, and visualization of ODDs.

Fraunhofer sets standards for AI in rail

The results of safe.trAIn were not only presented in scientific publications and conferences, but also submitted to standardization committees. Fraunhofer thus made significant contributions to the development of the DIN DKE SPEC 99002 standard DIN DKE SPEC 99002, which defines basic terms for AI in railway applications.

Particular emphasis was placed on the validation and possible approval of AI functions in rail systems. Finally, a conceptual assessment was also prepared by the partners of the participating TÜVs. This further shows how the developed safety concept and safety evidences can serve as a basis for future AI-based functions in rail.

safe.trAIn has demonstrated how trustworthy AI functions for safety-critical applications in railway can be developed and validated. The concepts, methods, and tools developed form a transferable basis for future applications – not only for rail transport, but also in other safety-critical domains such as mobility applications or industrial automation.


This work was funded by the European Union and the German Federal Ministry for Economic Affairs and Energy as part of the safe.trAIn project.

Read next

Artificial Intelligence
Railway AI Systems: The Importance of Operational Design Domain (ODD)

Christian Drabek / Fraunhofer IKS
Christian Drabek
Artificial intelligence / Fraunhofer IKS
Artificial intelligence