During the course of the KI Delta Learning project at the University of Applied Sciences Reutlingen, researchers addressed the challenge of domain shift in deep learning models used for autonomous systems. The team developed an unsupervised deep domain adaptation framework based on CycleGAN to transfer knowledge between synthetic and real sensor data without requiring labeled target data. The approach treats the source model as a black‑box regressor and learns a mapping that aligns the feature distributions of the two domains. The framework was evaluated on the demanding task of 2‑D human pose estimation, a key component for safety in autonomous driving and collaborative robotics. While the report does not report quantitative accuracy figures, the method was applied to a reference pose‑estimation network trained on synthetic imagery and subsequently adapted to real camera streams captured in the university’s motion‑capture laboratory. The adaptation pipeline was implemented using the open‑source Transfer‑Learning library by Jiang et al. and leveraged OpenCV for camera calibration and a custom simulation engine for synthetic data generation.

In parallel, the project produced a high‑quality dataset that underpins the domain‑adaptation experiments. In the AP1.2 work package, synchronized recordings were made with a Vicon motion‑capture system, Intel RealSense D435 and L515 cameras, and depth sensors. The motion‑capture rig automatically generated 2‑D and 3‑D pose annotations as well as 2‑D bounding boxes for pedestrians and micro‑mobility devices such as e‑scooters and hoverboards. The resulting corpus contains approximately 47,000 samples, comprising 45,620 image frames and 1,396 depth frames, all time‑stamped for perfect alignment. This dataset not only satisfies the project’s data‑collection goal but also provides a realistic testbed for evaluating robustness against corner cases, such as the presence of unfamiliar micro‑mobility vehicles that can cause overconfident predictions in standard models.

The project was carried out entirely within the university, with no external partners. Funding was provided by the German Federal Ministry of Economics and Technology under the “New Vehicle and System Technologies” program, which supports research on emerging automotive and mobility technologies. The research team’s responsibilities included data acquisition, synthetic data generation, development of the domain‑adaptation algorithm, and evaluation on pose‑estimation tasks. The outcomes of the project feed into subsequent initiatives such as the HEIDI project, which explores human‑technology interaction interfaces, and the AIDA project, which builds a large real‑world laboratory to blur the line between simulation and reality. The findings are also slated for dissemination to local industry partners.

The project’s scientific contribution was formally presented at the 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), where the paper “Auxiliary Task‑Guided CycleGAN for Black‑Box Model Domain Adaptation” was published. This work demonstrates a general, task‑agnostic approach to black‑box domain adaptation for regression problems, extending the state of the art beyond classification‑only methods. The dataset and the domain‑adaptation framework established by the KI Delta Learning project provide a solid foundation for future research and industrial applications in autonomous systems that must operate reliably across diverse real‑world conditions.