Institution profile

Mercedes-Benz AG

Industry researcheurope · de
Official website
Research library65linked papers
Opportunities0open roles
Selected work

Representative Papers

Carbon Footprint Evaluation of Code Generation through LLM as a Service

Mar 30, 2025

As AI code generation is increasingly deployed in high-reliability domains such as automotive systems, quantifying its embodied carbon (from development) and operational carbon (from execution) has become critically urgent. Method: This paper introduces the first code-level, full-lifecycle carbon footprint assessment framework tailored for LLM-based coding services—exemplified by GitHub Copilot—integrating hardware-aware and software-aware carbon modeling with established software sustainability metrics to yield a reproducible empirical evaluation pipeline. Contribution/Results: We demonstrate that carbon impact varies significantly across usage scenarios; moreover, green coding strategies substantially reduce functional carbon intensity (e.g., gCO₂e per feature or per executed line). This work delivers the first measurable, verifiable carbon assessment methodology for AI-generated code in safety-critical domains, enabling evidence-based green AI development practices and informing sustainable AI policy formulation.

1 citationsRead paper

FAM-HRI: Foundation-Model Assisted Multi-Modal Human-Robot Interaction Combining Gaze and Speech

Mar 11, 2025arXiv.org

Existing human-robot interaction (HRI) systems rely on single-modal inputs—such as gesture or speech—leading to high ambiguity, low efficiency, and poor accessibility for users with motor impairments. To address these limitations, this work proposes a lightweight gaze-speech bimodal interaction framework tailored for users with physical mobility constraints. Leveraging Meta ARIA smart glasses, the system captures real-time eye-tracking and speech signals, integrating temporal eye-movement modeling, adaptive gaze-duration estimation, vision-language alignment, and large language model (LLM)-driven intent parsing with contextual scene injection—all executed on-device. This establishes the first on-device LLM-powered real-time HRI paradigm, effectively suppressing ocular noise. Evaluation demonstrates >92% task success rate and sub-2.1-second average interaction latency. The framework exhibits exceptional robustness and usability in trials with physically impaired users.

1 citationsRead paper
Recent publications

Latest Papers

Drive-to-Music: Context-Aware Generative Audio for In-Vehicle Experiences

Aug 12, 2026

This work proposes a real-time, context-aware music generation method leveraging multimodal driving signals—such as dashcam imagery and vehicle telemetry—to enhance driver experience, attention, and well-being. By jointly modeling scene semantics and driving context, the system maps dynamic driving states to high-level musical attributes, which condition a generative audio model to synthesize low-latency, context-aligned soundtracks. The approach establishes the first end-to-end mapping from visual and motion inputs to structured musical features, enabling smooth transitions in response to evolving driving conditions. Integrated constraint-based controls and safety mechanisms ensure reliable deployment in automotive environments. Experimental results demonstrate the feasibility of real-time, adaptive audio generation within authentic in-vehicle settings, laying the groundwork for personalized intelligent cabin experiences.

0 citationsRead paper