Scholarlink

<p>To implement software for autonomous driving in Formula Student competitions, frequent retraining of neural networks (NNs) is critical to adapt to evolving datasets and dynamic team structures. This paper presents a Continuous Integration (CI) pipeline that automates data labeling, augmentation, training, and validation, focusing on lightweight NNs for perception tasks. Initially developed for Sensor-Fusion within the perception component, the pipeline was later adapted to the Vision-Only approach. The challenges of transitioning pipelines across differing sensor setups and software solutions and the feasibility of reusing existing artifacts are evaluated, highlighting the resilience and limitations of CI pipelines over time. Insights from deployment emphasize the importance of standardization and robust documentation to ensure adaptability and sustainability in rapidly evolving software environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evolution of the continuous integration of neural networks in autonomous systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年2月1日

查看全文
  • Bruno Steffen,
  • Leonard Sondermann,
  • Lukas Schütz

摘要

To implement software for autonomous driving in Formula Student competitions, frequent retraining of neural networks (NNs) is critical to adapt to evolving datasets and dynamic team structures. This paper presents a Continuous Integration (CI) pipeline that automates data labeling, augmentation, training, and validation, focusing on lightweight NNs for perception tasks. Initially developed for Sensor-Fusion within the perception component, the pipeline was later adapted to the Vision-Only approach. The challenges of transitioning pipelines across differing sensor setups and software solutions and the feasibility of reusing existing artifacts are evaluated, highlighting the resilience and limitations of CI pipelines over time. Insights from deployment emphasize the importance of standardization and robust documentation to ensure adaptability and sustainability in rapidly evolving software environments.