Automated Repair of Cyber-Physical Systems

TL;DR

This paper investigates automated repair of CPS software by addressing limitations in fault localization, test-time, and evaluation within CPS contexts. The proposed framework blends time-aware spectrum-based fault localization with an archive-based search that uses LLMs as patch mutators, guided by CPS-specific fault diagnostics and test data. It also includes a patch validation mechanism to curb LLM hallucinations and a unified debugging approach to handle stubborn bugs, validated on open-source and industrial CPS benchmarks. The anticipated contributions include a time-aware fault localization method, an open-source repair engine, and a CPS fault benchmark enabling reproducible evaluation, with potential to significantly reduce CPS debugging costs.

Abstract

Cyber-Physical Systems (CPS) integrate digital technologies with physical processes and are common in different domains and industries, such as robotic systems, autonomous vehicles or satellites. Debugging and verification of CPS software consumes much of the development budget as it is often purely manual. To speed up this process, Automated Program Repair (APR) has been targeted for a long time. Although there have been advances in software APR and CPS verification techniques, research specifically on APR for CPSs is limited. This Ph.D. research project aims to develop scalable APR techniques for CPSs, addressing problems of fault localization, long test execution times, and fitness function limitations. A new method combining spectrum-based fault localization (SBFL) with patch generation and advanced artificial intelligence techniques will be investigated. The approach will be validated by empirical studies on open and industrial code bases of CPSs.