Technical Summary: Foundation Models for Software Engineering of Cyber-Physical Systems: The Road Ahead

Problem Statement

The integration of Foundation Models (FMs), particularly Large Language Models (LLMs), into Software Engineering (SE) is rapidly expanding. However, their application to the Software Engineering of Cyber-Physical Systems (CPSs) remains underexplored and limited. Existing research predominantly focuses on LLMs for general SE tasks or the operational use of FMs within CPSs (e.g., for robot control), neglecting the broader software development lifecycle.

CPSs present unique challenges that distinguish them from conventional software systems. They are multidisciplinary, integrating hardware, software, communication, and physical environments. They process diverse, multimodal data (images, audio, sensor streams, text) and operate under strict real-time, safety, and reliability constraints. Current SE practices for CPSs struggle with:

• Data Heterogeneity: Handling multimodal inputs and outputs across the lifecycle.
• Safety and Certification: The difficulty of qualifying FM-generated artifacts for safety-critical standards.
• Domain Specificity: The lack of large, open corpora for proprietary CPS languages (e.g., PLC code, Simulink).
• Reality Gaps: The disconnect between simulation-based testing and real-world physical behavior.
• Uncertainty: Managing uncertainties arising from environmental conditions, sensor inaccuracies, and the non-deterministic nature of FMs themselves.

The paper argues that relying solely on text-based LLMs is insufficient. There is a critical need to explore other FMs, such as Vision-Language Models (VLMs) and multimodal models, to support the full spectrum of CPS software engineering phases.

Methodology

This paper does not follow a systematic literature review (SLR) protocol. Instead, it adopts a visionary, forward-looking research roadmap approach. The methodology for developing the roadmap involved an iterative process:

1. Initial Discussion & Brainstorming: Authors used a hybrid of the Delphi method and the Nominal Group Technique. Authors independently identified challenges and opportunities based on their expertise and literature review, followed by structured group sessions to reach consensus.
2. Internal Validation: Senior authors cross-validated sections for clarity, consistency, and correctness.
3. External Discussion: The roadmap was presented and discussed at the SE2030 Workshop (co-located with FSE 2025) using "liberating structures" (e.g., flash keynotes, fishbowl) to gather feedback from diverse external participants.
4. Integration & Extension: Authors incorporated workshop feedback, revised arguments, and extended the content.
5. Final Validation: A final comprehensive review was conducted before submission.

The roadmap is structured around six core software engineering phases and six cross-cutting dimensions. For each identified challenge, the authors define actionable research opportunities framed as key research questions, accompanied by suggested methodologies, baseline approaches, and evaluation metrics.

Key Contributions

1. A Comprehensive Research Roadmap

The paper presents the first systematic effort to offer a high-level, visionary guide for integrating FMs into the entire CPS software engineering lifecycle. It moves beyond LLMs to include VLMs and multimodal models, addressing the specific data modalities inherent to CPSs.

2. Phase-Specific Challenges and Opportunities

The roadmap details challenges and research opportunities across six phases:

• Requirements Engineering: Focuses on handling ambiguity in natural language requirements, verifying multimodal artifacts (text, diagrams, sensor data), identifying uncertainties, and establishing traceability. It proposes using multimodal FMs to translate requirements into formal models and vice versa.
• System Design and Modeling: Addresses the generation of physics-grounded models, handling heterogeneous inputs, and creating Digital Twins (DTs). It suggests using FMs for interactive modeling assistants, end-to-end model generation from natural language, and fine-tuning FMs to act as live DTs.
• Software Development: Highlights the challenge of generating code for proprietary, domain-specific languages and the need for safety-critical code qualification. It proposes lightweight, on-premise FMs for edge deployment and hybrid code review pipelines.
• Testing: Tackles the "reality gap" between simulation and physical testing. It suggests using FMs to generate high-fidelity synthetic test data, bridge the sim-to-real gap, and act as test oracles or surrogate implementations for differential testing.
• Debugging and Repair: Focuses on fault localization in complex software-hardware-physical interactions. It proposes multi-modal temporal fault localization and using FMs as surrogate test oracles to reduce the cost of expensive system-level testing.
• Evolution: Discusses proactive evolution, self-healing mechanisms, and the automated evolution of Digital Twins to reflect changes in the physical CPS.

3. Cross-Cutting Dimensions

The paper identifies and analyzes challenges across six dimensions that span all phases:

• Technical: Hallucinations, prompt sensitivity, and real-time inference constraints.
• Safety & Certification: The lack of verifiable FM behaviors and incompatibility with existing safety standards (e.g., ISO 26262, DO-178C).
• Economic & Resource: High costs of training, fine-tuning, and inference, especially for edge devices.
• Human, Organizational & Social: Overreliance on FMs, communication gaps in multidisciplinary teams, and intrinsic biases.
• Ethical & Privacy: Data leakage, intellectual property ownership of generated artifacts, and ethical biases.
• Environmental: The substantial carbon footprint of FM training and inference.

4. Actionable Research Framework

For every identified challenge, the paper provides a structured framework for future research, including:

• Research Questions: Specific inquiries to guide investigation.
• Methodologies: Suggested technical approaches (e.g., fine-tuning, prompt engineering, hybrid neuro-symbolic approaches).
• Baselines: Existing methods to compare against (e.g., manual engineering, traditional search-based testing).
• Evaluation Metrics: Quantitative and qualitative measures to assess cost-effectiveness and performance (e.g., fault detection rates, reduction in manual effort, simulation fidelity).

Results and Claims

As a research roadmap, this paper does not present empirical results, experimental data, or validated case studies. Its "results" are the identification and structuring of the research landscape.

The paper claims that:

• Current FM adoption in CPS SE is limited and overly focused on LLMs, ignoring the potential of multimodal models which are better suited for the diverse data types of CPSs.
• Significant research gaps exist in integrating FMs into safety-critical workflows, particularly regarding certification, real-time constraints, and the "reality gap" in testing.
• A structured, phase-based approach is necessary to make FM integration accessible to software engineers with diverse expertise.
• The proposed roadmap serves as a foundational guide to direct future empirical studies and to bridge the gap between general FM capabilities and the rigorous demands of CPS software engineering.

Significance

The paper positions itself as a visionary guide rather than a definitive solution. Its significance lies in:

1. Broadening the Scope: It shifts the focus from LLMs alone to a broader class of Foundation Models (including VLMs and multimodal models) essential for CPSs.
2. Systematizing the Field: It provides the first comprehensive map of challenges and opportunities across the entire CPS software lifecycle, moving beyond isolated studies on specific tasks.
3. Actionability: By providing concrete research questions, methodologies, and metrics, it transforms high-level concepts into actionable research agendas for the community.
4. Holistic View: It explicitly addresses non-technical barriers (economic, ethical, environmental, and human factors) that are often overlooked in technical roadmaps but are critical for the real-world adoption of FMs in safety-critical domains.

The authors emphasize that the roadmap is intended to be iteratively refined as new studies and FM paradigms emerge, serving as a living document to guide the future of CPS software engineering.