2026

Abstract: Mapless driving lacks globally consistent lane-level route information from an HD map. At the same time, motion planning still depends on local lane structure produced online by the perception and mapping stack. We address this gap by conditioning planning on natural-language route descriptions that encode global routing intent and aligning them with learned embeddings of the local map. To study this setting at scale, we extend nuPlan with 60-second natural-language route descriptions for 1 000 000 training scenarios and the Val14 split, train language-conditioned models on this extension, and evaluate on Val14. Compared to route-conditioned baselines, our language-conditioned models achieve a better trade-off between driving quality and route compliance, with particularly strong gains in long-horizon navigation compliance. In our experiments, we use ground-truth lane-level local maps in place of online mapping outputs to isolate the effect of route-language conditioning. The dataset is available at github.com/KIT-MRT/nuplan-language-based-routing-dataset.

Abstract: Data-driven predictive control (DPC) optimises a system’s behaviour directly from measured input–output trajectories without relying on an explicit state-space model. This property is attractive for autonomous driving, where accurate models can be difficult to obtain across operating conditions. In practice, however, more data improves prediction quality at the cost of a larger optimisation problem, creating a tension between tracking accuracy and real-time feasibility. We complement the standard DPC formulation with a set of practical design choices that substantially affect numerical robustness and closed-loop performance in practice. We apply DPC to trajectory tracking on a scaled autonomous vehicle and investigate the trade-off between dataset size, tracking accuracy, and computational effort by varying the number of trajectories through random sampling. A model-based MPC controller using a known kinematic bicycle model serves as a baseline, providing context for the achievable tracking accuracy and computation time.

Abstract: Recent research in machine and deep learning has shown the potential of learningbased motion planning approaches to improve the driving behavior of automated vehicles, especially in complex environments. However, their complex nature and lack of transparency can hinder explainability and trustworthiness and complicate safety assurance. Motivated by these challenges, we propose a hybrid planning architecture that combines the advantages of machine learning with the verifiability and the determinism of classical approaches. Specifically, we developed a deep neural network to interpret complex traffic scenes and propose driving behavior, while an optimization-based supervision layer validates this proposal and enforces explicit drivability and safety constraints. We evaluate the learned planner’s driving behavior in open-loop studies on real-world urban data, discuss system integration aspects for stable closed-loop operation, and report results from real-world deployment on our research vehicle karl..

Abstract: Deep learning for automotive radar perception relies on large labeled datasets. Such data are difficult to obtain, as access to raw radar signals is limited and realistic simulation requires substantial effort. This paper presents a cross-radar supervision framework for automatic labeling of 4D automotive multiple-input multiple-output (MIMO) radar data using synchronized radar sensors. Two evaluation radars with different antenna configurations are operated alongside an industrialized high-resolution production radar, whose detections provide reference labels. A costfunction-based post-processing method associates production-radar detections with evaluation-radar measurements in the range-Doppler domain. Qualitative results from real-world driving scenarios show consistent range-Doppler structures across the evaluation radars and plausible projection of production-radar detections into range-Doppler and spatial representations. The generated labels are intended as reference labels for weak supervision rather than bin-exact ground truth. The proposed framework therefore provides a practical route toward automatically labeled 4D radar datasets while preserving access to evaluation-radar data across multiple processing stages, from raw radar measurements to range-Doppler, spatial, and detection-level representations.

Abstract: Monocular 3D object detection offers a scalable perception solution for autonomous driving, yet models trained on perfectly annotated synthetic data suffer performance drops in real-world environments due to the severe sim-to-real domain gap. We study a weakly supervised domain adaptation setting in which the synthetic source domain provides full 3D annotations, while the real target domain provides only auxiliary 2D bounding-box annotations and no target-domain 3D labels for detector supervision. Mean Teacher self-training attempts to bridge this gap but frequently falls into a “confidence trap.” Confronted with unfamiliar real-world data, the teacher model generates noisy pseudo-labels, causing standard confidence-thresholding methods to aggressively discard predictions and ultimately deprive the student model of a viable supervisory signal. To overcome this, we propose a novel Weakly Supervised Domain Adaptation (WSDA) framework that integrates off-the-shelf foundation models into the self-training loop to provide external object-level guidance. By prompting Segment Anything v2 (SAM2) with target-domain 2D box centers to generate car-candidate object masks, we independently validate the teacher’s 3D bounding box predictions against mask consistency criteria, filtering out mask-inconsistent False Positives. Furthermore, we introduce a False Negative (FN)-Whitening strategy that explicitly erases the visual features of missed objects in the input image. This creates a don’t-care-like input masking region that preserves surrounding scene context while reducing destructive background associations. Our mask-based filtering mechanism stabilizes the teacher-student dynamic in our experiments and mitigates the confirmation bias cycle. Extensive evaluations on our custom KITTI split show that the proposed framework improves over a standard self-training baseline, achieving a relative performance gain of over 34% in Moderate AP3D.

Abstract: Learning metric correspondences across heterogeneous sensor observations is a key requirement for robust Bird’s-Eye View (BEV) perception in autonomous driving. This paper studies uni-, cross-, and multi-modal keypoint matching as a unified latent-space problem, where camera and LiDAR observations are projected into a common BEV representation and matched through modality-agnostic keypoints and descriptors. We instantiate this formulation through BEVMatch, a self-supervised framework that learns metric keypoints and descriptors directly in BEV space.
The framework employs modality-specific encoders, a cross-modal fusion module, and lightweight prediction heads operating on a shared BEV latent space. Training uses relative odometry as geometric supervision and modality dropout to encourage sensor-agnostic representations without requiring manual keypoint annotations.
Experiments on nuScenes demonstrate that a single model can estimate correspondences across camera-only, LiDAR-only, cross-modal, and multi-modal settings, including scenarios with larger relative translation and rotation. To assess practical applicability, we further integrate the learned correspondences into a keypoint-based SLAM pipeline deployed on a research vehicle. While the resulting system is evaluated as a feasibility study rather than a productionready localization solution, the results indicate that unified metric BEV correspondences can support downstream localization and mapping tasks across varying sensor configurations.

Abstract: End-to-end learning is widely used in autonomous driving, yet its stability under limited perception is not well understood. We analyze end-to-end driving policies operating with limited visual resolution and sparse two-dimensional range sensing. Using a controlled simulation setup, we evaluate lightweight imitation-learned architectures and assess their behavior in closedloop driving. Our results reveal characteristic failure modes such as accumulated drift and oscillatory control, which are influenced by control dimensionality, state feedback, and sensor fusion. These insights clarify the conditions under which end-to-end approaches remain stable and highlight directions for enhancing robustness through temporal modeling, structural priors, and integration of additional constraints.

Abstract: External human-machine interfaces (eHMIs) are widely proposed to replace the implicit cues that disappear with driverless autonomous vehicles (AVs). Vulnerable road user (VRU) diversity, including but not limited to age, mobility, sensory capability, language, culture and emotional state, primarily determines eHMI effectiveness. Attention-reducing factors systematically affect pedestrian crossing behaviour, while none of the established AV simulators or VR laboratory setups simultaneously provide traffic dynamics, configurable eHMIs and VRU diversity. This paper, therefore, proposes a simulation prototype that targets the human-facing aspect of the AV-to-VRU communication.

Abstract: Previous research in the context of autonomous bus shuttles has focused mainly on technical and safety-related acceptance determinants while the role of human interaction in the use of autonomous public transport systems has received comparatively little attention. The research project IMoGer (Innovative modular Mobility Made in Germany) aims to investigate conditions for the implementation of autonomous mobility services based on modular electric vehicles. In this context, societal readiness lays an important role. A quantitative user study (N = 283) is presented that focused on the relationship between individuals’ need for interaction with driving personnel and their intention to use autonomous bus shuttles. The results will be used for further development of the autonomous vehicle technology and service concept in the project and to derive recommendations for researchers and practitioners.

Abstract: Automated-driving user experience depends not only on whether a maneuver succeeds, but also on how its motion unfolds. This exploratory follow-up study tests whether in-cabin facial emotion-channel probabilities co-vary with vehicle dynamics during Level-2 automated urban turns, which motion cues dominate, and which associations remain after participant-baseline control. Across 44 participants, 497 clips, and 2626 two-second windows, weak but reproducible coupling emerged. Lateral acceleration and speed were most consistent, with yaw/curvature and longitudinal jerk secondary. Facial channels are therefore treated as contextual behavioral markers, not standalone emotion detectors.

Abstract: Virtual scenario-based testing is a key enabler for validating automated driving systems (ADS) and intelligent transport systems (ITS). However, executing large-scale test suites involving possibly thousands of scenarios remains labor-intensive and difficult to scale. This paper presents an end-to-end, DevOps-driven framework that automates build, deployment, and distributed execution of CARLA-based scenario tests of an ADS on a lightweight Kubernetes cluster. ROS 2 applications are packaged as standardized Kubernetes Helm charts generated from repository specifications, while entire simulation environments are composed declaratively via dynamic Helmfile manifests. The paper describes how a distributed testing workflow can be implemented in Argo Workflows to provision environments, aggregate and batch OpenSCENARIO test cases from configurable sources, execute scenarios in parallel across cluster nodes, and collect logs and resource metrics. In an evaluation on a multi-node K3s cluster running 200 scenarios, the best configuration speeds up end-to-end workflow time by more than a factor of eight compared to a sequential baseline. The results demonstrate significant gains in end-to-end execution time and quantify trade-offs between parallelism, orchestration overhead, and cluster stability. The framework is further demonstrated in a real-world ADS test application with connections to scenario sources and downstream evaluation modules. This demonstrates that the approach provides a strong foundation not only for scalable simulation testing, but also for generating traceable evidence that can support safety arguments.

Abstract: Artificial Intelligence (AI) represents a key enabling technology for automated driving functions, particularly in perception and planning tasks. However, ensuring the safety of AI-based systems remains challenging. Key reasons include black-box characteristics, limitations in robustness and reliability, and possible deviations from the intended scope of application. Given the difficulties in exhaustively testing these systems, the need for monitoring during operation after deployment is of high importance. To address these challenges, the recently published ISO/PAS 8800 standard introduces a structured framework for the development and operation of AI-based systems in safety-related automotive contexts. While the standard defines an AI safety lifecycle and related measures, it leaves considerable freedom in their technical realization. Runtime monitoring is a key mechanism within this context, enabling the detection of safety-relevant deviations during operation. However, the diversity of existing monitoring approaches leads to challenges for their systematic comparison, selection, and integration. This paper provides a comprehensive overview of existing monitoring taxonomies and proposes a novel taxonomy for AI monitors aligned with ISO/PAS 8800. The proposed taxonomy aims to improve the traceability and comparability of monitoring strategies, thereby strengthening AI safety argumentation and aiding in the fulfillment of relevant safety standards.

Abstract: Well-designed autonomous driving datasets have fundamentally shaped research progress, yet existing literature primarily describes what datasets contain rather than how to strategically design impactful ones. This is especially limiting for small and medium-sized labs and startups that cannot afford to misallocate scarce resources. We argue that impactful dataset creation begins with a diagnosis: whether a research question is blocked by a data problem or an evaluation problem, and proceeds by selecting the minimal data operator(s) that closes the resulting gap, recording new data only when no cheaper operator(s) suffices. We analyze the evolution of major autonomous driving (AD) datasets through this lens and distill a strategic framework spanning gap identification, operator choice, sensor suite design, and annotation strategy. We ground the framework in a running case study of our KITScenes dataset family. The datasets are available at: https://kitscenes.com/.

Abstract: Urban environments present significant challenges for the perception systems of automated driving due to their high complexity, characterized by dense traffic, diverse obstacles, or a wide range of possible scenarios. While numerous automotive datasets exist, few focus on urban settings, and the majority are acquired exclusively on passenger cars and overlook other transport modalities such as heavy-duty trucks, busses and trams. In contrast to passenger vehicles, these platforms operate under different geographical and operational constraints; including traversing pedestrian zones or platform-specific tasks such as movement of trailer or winter service. In this work a method is proposed for a guided dataset creation based on an analysis of existing datasets using user-generated map data. Furthermore, representative truckspecific use cases are identified and exemplified to demonstrate the applicability of the proposed methodology and to support targeted acquisition planning.

Abstract: “Safety” and “risk” are key concepts in the context of automated vehicle deployment. A clear conceptualization of both concepts is not only necessary for safety engineers who are responsible for assurance processes. For the communication with and between technical and non-technical stakeholders (engineers, regulators, lawyers, the general public) in particular, common, or at least compatible, notions of these abstract concepts are crucial for effective communication and for setting mutual expectations.
In the European market, automated vehicles require type approval or test permits which are subject to current regulatory efforts. The resulting regulatory documents are a central means of communication between regulators and the engineers designing and developing the technical systems. Flawed terminology regarding regulatory safety expectations for automated vehicles can unnecessarily complicate relations between regulators and manufacturers and thus hinder the introduction of the technology. In this paper, we set a focus on the European market and review relevant documents at the UN- and EU-level, for the UK, and Germany regarding their (implied or explicit) notions of safety and risk. We map regulatory notions to established as well as more recently developed notions of safety and risk in the field of automated driving, pointing out potential conflicts. Based on the analysis, we provide requirements and recommendations for establishing clear definitions of safety and risk in regulations. Furthermore, we discuss how applying such clear definitions and acknowledging their consequences can support rather than hinder the market introduction of automated vehicles.

Abstract: Although automated driving systems (ADS) increasingly operate in public traffic, cooperation between automated vehicles and humans must not be forgotten in light of improving safety and efficiency. However, deploying such cooperative functions is challenging in real traffic situations. This paper presents developments from the German National project Managed Automated Driving (MAD) Urban using the open-source Eclipse-Automated Driving Open Research (ADORe® ) stack based on ROS2. ADORe® enables coordinated decision making through two complementary modes: Single-Agent Autonomous Driving (SAAD) and Multi-Agent Autonomous Driving (MAAD). While ADORe SAAD provides robust onboard automated driving, capable of Vehicle-to-Everything (V2X) cooperation and remote operations, ADORe MAAD is the version designed to run on the infrastructure, allowing to jointly predict and plan trajectories for multiple vehicles, by using infrastructure sensing and Infrastructure-to-Everything (I2X) communication. Together, these components form a cooperative driving framework designed to function reliably under realistic conditions, which have been deployed in real traffic situations.