RESHAPE: Requirements Engineering for System–model–data Hierarchy and Asymmetric V-Process in E2EAD

Companion repository for the paper: "RESHAPE: Requirements Engineering for System–model–data Hierarchy and Asymmetric V-Process in End-to-end autonomous driving" — Under review

Hanyang University · Department of Automotive Engineering
Hyundai Motor Company · Autonomous Driving System Development Team

---

Overview

RESHAPE addresses the requirements engineering gap in end-to-end autonomous driving (E2EAD), where conventional RE processes provide no structure for specifying model and data requirements or tracing their relationships to system-level safety goals.

The work makes three contributions corresponding to three research questions:

RQ	Contribution	Output
RQ1	LLM-assisted Systematic Mapping Study identifying 17 requirement element groups across System, Model, Data, and cross-layer concerns	Element catalog (S1–S7, M1–M4, D1–D4, CL1–CL2)
RQ2	SMDL V-Process — a 12-step process assigning each layer an independent V-cycle with top-down specification and bottom-up verification	Process model with asymmetric execution order
RQ3	LLM-based multi-agent authoring framework with Expert–Auditor dyads and cross-layer traceability orchestration	SRS, MRS, DRS, RTM, Review Register

Figure 1. Overview of the SMDL V-Process. Specification proceeds top-down (System → Model → Data); realization proceeds bottom-up (Data → Model → System).

Figure 2. Multi-agent requirement authoring framework. Pre-knowledge curation from the SMS element catalog and ISO standards, followed by requirement authoring via Expert–Auditor dyads, Orchestrator integrity checks, and Formalization Agent.

Key Concepts

• SMDL (System–Model–Data Layer): Three-layer hierarchy where system requirements are refined top-down into model and data requirements, while verification results propagate bottom-up
• Asymmetric V-Process: Specification proceeds System → Model → Data; realization and verification proceed Data → Model → System, driven by physical dependency (data must exist before training)
• Expert–Auditor dyads: Each layer is served by a generation agent (Expert) and a review agent (Auditor), structurally suppressing self-confirmation bias
• Requirement Draft Units (RDUs): Atomic specification records combining EARS-formatted content, acceptance criteria, inter-layer traceability links (traced_to / derived_from), and approval status
• Cross-layer concerns: CL1 (Open-loop ↔ Closed-loop Gap) and CL2 (Data–Model–System Traceability) — structural challenges that span multiple layers
• Plausible hallucination: LLM-generated values that appear authoritative but lack grounding in provided input — mitigated by Auditor discipline and review_required marking

Evaluation

RQ2: Retrospective Gap Analysis

Ten real development problem instances from collaborative E2EAD development with Hyundai Motor Company, analyzed against the SMDL element catalog. Seven of ten could have been prevented through upfront multi-layer specification.

RQ1+RQ2+RQ3: Specification Quality Comparison

Two-condition comparison (Baseline vs. RESHAPE) evaluated by 10 practitioners from Hyundai Motor Company and an LLM evaluator (Claude Opus 4.6), using 11 criteria across four tiers (statement quality, document quality, epistemic integrity, practical usefulness).

Compared Configurations

Both configurations use the same base LLM (Claude Sonnet 4.5), RDU schema, Orchestrator, and Formalization Agent. The two differences are: (i) domain knowledge depth (element names only vs. standards-grounded pre-knowledge), and (ii) the presence of the Auditor.

	Baseline (Expert-only)	RESHAPE (Expert + Auditor)
Pre-knowledge	Element names only	SMS-derived scope descriptions + ISO standard clauses via RAG
Review	None	Auditor cross-checks against pre-knowledge
Uncertainty marking	Values presented as approved facts	Ungrounded values marked review_required / TBD
Reflection iterations	None	Up to K=3 per dyad

Results

	Baseline	RESHAPE
RQ2: Gap Analysis	—	7/10 problems preventable
Practitioner Score (3-point scale)	2.21	2.74
> Epistemic integrity (C.1 Source traceability)	1.4	2.6
> Epistemic integrity (C.2 Uncertainty marking)	1.8	2.9
> Practical usefulness (D.1–D.3 mean)	2.07	2.77

Supplementary Materials

Material	Location	Description
SMS protocol	systematic_mapping_study/SMS_Protocol.md	Full procedural details: inclusion/exclusion criteria, keyword rationale, search query, 9-step procedure, extraction/clustering rules
SMS agent prompts	systematic_mapping_study/SMS_Agent_Prompts.md	Prompts used for LLM-assisted literature collection and snowballing
SMS results	systematic_mapping_study/result/	Corpus, extraction labels, and mapping data
Element scope descriptions	smdl_framework/Requirement_Element_Scope.md	Per-element scope descriptions for S1–S7, M1–M4, D1–D4, CL1–CL2
RDU schema	smdl_framework/Requirement_Draft_Unit_Schema.md	Complete RDU field list and worked examples
Gap analysis details	evaluation/RQ2_Retrospective_Gap_Analysis.md	Detailed problem instance descriptions mapped to SMDL elements
Evaluation rubric	evaluation/RQ3_Evaluation_Rubric.md	11-criteria scoring rubric (Tiers A/B/C/D) used for specification quality assessment

Citation

@inproceedings{na2026reshape,
    title     = {{RESHAPE: Requirements Engineering for System--model--data Hierarchy and Asymmetric V-Process in End-to-end autonomous driving}},
    author    = {Na, Yuseung and Kim, Hyunjun and Kim, Donggue and Ryu, Seungji and Kim, Youngki and Bong, Sechang and Song, Saheon and Jo, Kichun},
    booktitle = {TBD},
    year      = {2026},
}

License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Acknowledgments

This work was supported in part by the Autonomous Driving System Development Team, Hyundai Motor Company and in part by the National Research Foundation of Korea (NRF) funded by the Korean Government Ministry of Science and ICT (MSIT) under grant No. RS-2024-00421129.