MAN Truck & Bus

Large Language Models (LLMs) can generate Computer-Aided Design (CAD), yet lack physical comprehension required for reliable engineering design. Instead of attempting to implicitly learn physical laws from data, we propose a Hybrid Agentic-Physical Architecture that embeds validated knowledge-based engineering tools directly into the decision-making loop of autonomous AI agents. In this framework, engineering design is formulated as a closed-loop, sequential decision-making process guided by explicit physical verification. Based on a load case, dedicated agents iteratively plan, generate, evaluate, and revise engineering designs using knowledge-based tools as a feedback signal. We introduce a benchmark dataset and metrics for assessing functional validity in generative CAD. Our system generates more complex and physically verified designs, with a 4.2x increase in structural complexity and improving compile rate by 3.5% compared to similar agentic methods. The codebase, prompts and dataset will be made publicly available to support reproducibility and future research.

Current generative methods for Computer-Aided Design (CAD) are narrowly specialized, requiring separate models for tasks such as Text-to-CAD or B-Rep-to-CAD translation. This fragmentation hinders useable CAD generation and limits deployment in engineering workflows. We hypothesize that a unified tokenization, combined with a decoder-only architecture, enables efficient cross-task transfer while reducing model complexity. We present the first multi-task CAD transformer that handles diverse sequence-to-sequence tasks within a single framework. Our unified embedding space reduces token count by $84.8 \%$ compared to existing approaches. Our compact $\mathbf{1 2 8 M}$-parameter decoder is pre-trained on a large-scale CAD dataset and adapted across tasks by adding lightweight cross-attention encoders. Our pre-trained model accelerates convergence by up to $\mathbf{5 4 . 8} \boldsymbol{\%}$ during task-specific fine-tuning and matches or exceeds the accuracy of models $50 x$ larger. Further, we are the first to leverage the strict syntax of CAD in our decoding strategy, reducing invalid generation rates from $\mathbf{4 . 4 \%}$ to $\mathbf{3 . 9 \%}$. This establishes a foundation for general-purpose CAD AI systems that adapt to new design modalities without architectural changes. We make the code available at github.com/TheEliasBe/pretrained-multi-task-cad.

Abstract Volatile Rahmenbedingungen, häufige Störungen und hohe Komplexität erhöhen den Resilienzdruck in industriellen Wertschöpfungsnetzwerken. Der Beitrag zeigt auf, wie generative KI (GenAI) in den drei Resilienzphasen der Antizipation, Reaktion und Erholung unterstützend zum Einsatz kommen kann. Auf Basis einer Literaturrecherche und explorativer Experteninterviews werden Potenziale, Anwendungsfelder und Umsetzungsbarrieren systematisiert sowie Handlungsempfehlungen für die Industrie abgeleitet.

In recent years, vehicles have become increasingly intelligent and context aware. Studies in the passenger car and digital sectors show that systems which adapt to different situations and users can improve usability and safety. To identify the potential of adaptive human-machine interfaces (HMI) in the truck sector, this work presents three exploratory focus groups conducted with truck drivers and industry experts, involving a total of N = 19 participants. Using qualitative content analysis, problem areas were identified to be addressed through adaptive functions. These problems included individual driver needs as well as complex menu structures and challenges in navigation and logistics. The proposed solutions, such as personalization, will be examined in more detail in further research. Furthermore, the study revealed challenges associated with the implementation of adaptive interfaces in trucks, particularly patronization through technology. Practical Relevance: Existing adaptive solutions predominantly concern passenger car and digital domains and fail to take into account the particular requirements of the commercial vehicle sector. A commercial vehicle is a complex workplace, where drivers often operate under time pressure. Therefore, these insights provide a foundation for the user-centered design of future adaptive display and control concepts in the commercial vehicle domain.

Starter-battery failures can cause costly and disruptive downtime in heavy-duty fleet operation. In this work, we study whether remotely collected engine-start measurements collected with vehicles manufactured by MAN Truck & Bus can support early warning of imminent starter-battery failures. We formulate the task as binary classification and compare single-start and sequential deep-learning models on two anonymised fleet cohorts, further evaluating early warning at increasing prediction offsets from the event of interest. The best single-start model achieves mean test AUCs of. 818 on Cohort A and. 864 on Cohort B, while sequential models remain competitive but do not surpass the strongest single-start baseline. Performance degrades as the prediction offset increases, as expected. Overall, the results show that operational engine-start signals contain useful predictive information and that a compact binary classification setup provides a strong baseline for deployment-oriented batteryfailure warning.