Technical Summary: Studying, Identifying, and Fixing Hidden Technical Debt in AI-Intensive Cyber-Physical Systems

Problem Statement

Artificial Intelligence (AI) components are increasingly integrated into Cyber-Physical Systems (CPSs), creating complex "AI-CPS" architectures found in autonomous vehicles, robotics, and healthcare. While Technical Debt (TD)—the long-term maintenance cost resulting from short-term design decisions—is well-studied in conventional software, the interaction between AI models, firmware, hardware, and mechanical components in AI-CPSs introduces unique challenges.

Existing TD taxonomies fail to fully capture debt arising from the specific interplay of software intelligence and the physical world. Key gaps include:

• Hidden Debt: Debt that is not self-admitted by developers (Self-Admitted TD) but exists in artifacts like sensor configurations, simulation fidelity, and hardware constraints.
• Heterogeneity: Debt spans diverse artifact types, including source code, ML model cards, firmware, and physical specifications.
• Complex Interactions: Specific challenges arise regarding how AI autonomously enacts hardware, real-time requirements, safety safeguards, and regulatory compliance.

The core research question is: What kind of specific technical debt occurs in the development of AI-CPSs, and how can it be automatically repaid?

Methodology

The research follows a multi-stage, mixed-method approach divided into three phases, as illustrated in the proposed framework (Fig. 1):

Phase 1: Empirical Characterization

The goal is to build a foundational understanding of AI-CPS TD through data mining and practitioner interviews.

• Repository Mining: The study analyzes Hugging Face and GitHub repositories to track the lifecycle of pre-trained models (selection, integration, maintenance) and identify TD patterns. It examines commits, file changes, documentation, and metadata.
• AI-CPS Specifics: Mining focuses on distinguishing AI-CPS-specific debt (e.g., sensor robustness, actuation, calibration, trajectory generation) from conventional or general ML debt.
• Interviews: Semi-structured interviews with developers and researchers investigate the real-world adoption of open-weight AI models in CPSs, focusing on integration challenges and current mitigation strategies.
• Outcome: A taxonomy of AI-CPS-specific TD and a catalog of developer challenges.

Phase 2: Automated Identification and Mitigation

Building on the Phase 1 taxonomy, this phase develops methods to detect "hidden" TD (unadmitted by developers) and propose remediation.

• Agentic Classification: An agentic approach utilizes metrics, static analysis warnings, and hardware-related information mined from repositories.
• Multi-Source Analysis: The system analyzes diverse evidence beyond code comments, including:
  • CPS Artifacts: Sensor/actuator descriptions, middleware settings, simulation files, safety constraints, and deployment scripts.
  • ML Artifacts: Model cards, training/inference scripts, datasets, preprocessing pipelines, and hyperparameters.
• Mitigation Strategies: Remediation actions are empirically derived by analyzing how developers resolve debt in existing repositories and incorporating practitioner feedback on model selection and compatibility.

Phase 3: Agentic AI-Based Governance

The final phase integrates previous outputs into an autonomous governance framework for continuous TD management.

• System Architecture (Fig. 2): The proposed system comprises:
  • TD Identification Agent: Classifies hidden debt against the taxonomy.
  • Solution Recommender: Proposes and potentially enacts remediation.
  • Impact Assessment Agent: Evaluates effects on maintainability, reliability, and deployability.
  • Knowledge Base: Stores the taxonomy, challenges, and remediation solutions.
  • Specialized Analyzers & Retrievers: Mine data from code, firmware, hardware specs, and external hubs (Hugging Face, Kaggle).
  • Orchestrator: Manages the workflow, monitors repository changes, and generates governance reports.
• Human-in-the-Loop: The system is semi-automated, interacting with developers via IDEs to validate classifications and suggestions, using feedback to refine the taxonomy and prompts.

Key Contributions

The research aims to deliver the following specific contributions:

1. Empirical Characterization: A systematic understanding of AI model ecosystem evolution and adoption based on repository analysis and interviews.
2. AI-CPS TD Taxonomy: A novel classification of technical debt specific to the intersection of AI and CPS, complementing existing taxonomies for conventional and ML systems.
3. Practitioner Insights: A detailed account of challenges faced by developers integrating AI models into AI-CPSs.
4. Automated Techniques: Specific strategies for identifying and fixing TD across diverse artifacts (code, models, hardware configs) rather than just source code.
5. Agentic Governance Framework: A system for continuous monitoring, impact assessment, and mitigation recommendation using agentic AI.

Current Status and Preliminary Results

The project is currently in its second year with the following initial findings:

• Repository Analysis: An empirical study of Hugging Face and GitHub revealed that fewer than 15% of the most-downloaded Hugging Face models have a linked development repository on GitHub. Furthermore, model cards often undergo only minor updates, limiting their utility as live documentation.
• Taxonomy Construction: Analysis of Self-Admitted TD (SATD) in 15 open-source AI-CPS repositories has initiated the construction of a specific TD taxonomy. Preliminary categories include sensor robustness, simulation simplification, safety constraints, middleware workarounds, and model integration issues.
• Interviews: Practitioner interviews are ongoing, with plans to publish findings in software engineering conferences.

Significance and Future Work

The significance of this work lies in "democratizing" AI-CPS development by providing structured approaches to manage the complex, often hidden, technical debt inherent in these systems. By moving beyond code-centric debt analysis to include hardware and AI model interactions, the research aims to improve the reliability and maintainability of critical autonomous systems.

Validation Plan:
The agentic AI solution will be validated through:

1. Academic Studies: Initial testing with students in project settings.
2. Real-World Demonstrators: Evaluation using the InnoGuard project's demonstrators, specifically a humanoid robot and a Leo Rover. These scenarios provide realistic environments involving sensing, actuation, navigation, and safety constraints to assess the framework's ability to identify debt across software, models, configurations, and hardware.

Dissemination:
Results are planned for publication in software engineering conferences and journals, with a focus on TD classification and remediation in the near term, followed by the agentic AI infrastructure in the final year.