At the recent RoboTUNN project meeting at Fraunhofer IPM and INATECH in Freiburg, all partners came together to coordinate the next technical steps toward automated tunnel inspection. For us at StatSoft, one topic was central: advancing our models for AI-based damage classification — a key component of the digital twin and future predictive maintenance approaches.
Improving Crack Classification
Several important technical decisions were made regarding damage detection. Going forward, all predictions from the YOLO model will be transferred in full to the central data environment, allowing StatSoft to apply its own filtering logic to determine which detections are relevant.
This is essential for reproducible crack classification and for assessing the criticality of each damage instance.
Another step forward concerns the training of the underlying detection models. Our partners at INATECH are evaluating, among other things, a rebalancing of the training data to reduce class imbalance, as well as alternative architectures such as SAM3. In parallel, we are assessing heuristic post-processing algorithms to mitigate the well-known issue of overly wide crack annotations — a challenge faced across the entire field.
Predictive Maintenance: Initial Modeling Approaches
During workshop sessions, several approaches for predicting the future development of structural damage were discussed.
Particularly promising is the use of experimental data from RWTH Aachen, where crack formation is recorded under controlled conditions. These datasets could provide valuable insights into how existing damage evolves between inspections. Additional ideas include moisture detection, monitoring joint displacements via repeated LiDAR scans, and integrating weather data.
All approaches will be evaluated by the end of January — an important step toward the first predictive maintenance prototype planned for autumn 2026.
Integration of Sensor Systems
At the same time, INATECH and RWTH are progressing with the integration of the multimodal sensor system onto the robots. This work is crucial for StatSoft, as image, thermography, and point cloud data will soon feed jointly into our classification and assessment models.
Conclusion
The meeting in Freiburg highlighted the strong momentum of the RoboTUNN project. For StatSoft, the resulting next steps are clear: more precise data pipelines, enhanced classification models, and deepened collaboration with our partners. The recent advances in crack classification, in particular, bring us closer to our goal of not only automating tunnel inspection processes but also making them predictable in the long term.
