RISC-V Functional Safety for Autonomous Automotive Systems: An Analytical Framework and Research Roadmap for ML-Assisted Certification N. Andreasyan∗ , M. Struve∗ , A. Popov† , M. Nikolaev† , and V. Vashkelis†
arXiv:2604.17391v1 [cs.SE] 19 Apr 2026
∗ Automotive Safety Lab † Embedded Intelligence Lab
{research,safety}@emilab.org Abstract RISC-V is emerging as a viable platform for automotive-grade embedded computing, with recent ISO 26262 ASIL-D certifications of commercial RISC-V processor IP demonstrating readiness for safety-critical deployment in autonomous driving systems. However, functional safety in automotive systems is fundamentally a certification problem rather than a processor problem: the dominant costs arise from diagnostic coverage analysis, toolchain qualification, fault injection campaigns, safetycase generation, and compliance with ISO 26262, ISO 21448 (SOTIF), and ISO/SAE 21434. This paper provides a structured analytical framework and expert-driven research roadmap for enabling economically superior certification of automotive-grade RISC-V platforms. Rather than proposing a single algorithmic breakthrough, we contribute: (1) a certification economics framework with formal cost and advantage models for strategic decision support; (2) an MLassisted certification framework mapping LLMs, knowledge graphs, reinforcement learning, and graph neural networks to specific certification bottlenecks; (3) a five-level RISC-V Safety Maturity Model (RSMM) for evaluating platform certifiability; and (4) an illustrative case study applying the framework to an ASIL-D autonomous emergency braking (AEB) ECU. We present a structured comparison of Arm and RISC-V architectures across eight functional safety dimensions, address the convergence of ISO 26262, ISO 21448, and ISO/SAE 21434 for autonomous driving, and define quantitative evaluation metrics for certification cost reduction. The paper is positioned as an analytical perspective contribution that structures the research space and identifies the highest-value problems for the RISC-V automotive certification ecosystem. Index Terms RISC-V, Functional Safety, ISO 26262, SOTIF, Autonomous Driving, ASIL-D, Certification Economics, Machine Learning, Safety Case Automation, FMEDA, Toolchain Qualification
I. I NTRODUCTION The automotive industry is undergoing a structural transformation from distributed Electronic Control Units (ECUs) toward centralized compute platforms and Software-Defined Vehicles (SDVs). Advanced Driver Assistance Systems (ADAS), Level3 and Level-4 autonomous driving, battery management, zonal architectures, and over-the-air (OTA) software updates are fundamentally changing the safety requirements of automotive systems [1]–[3]. Functional safety has consequently become a first-order engineering constraint. ISO 26262 defines the framework for functional safety of electrical and electronic systems in road vehicles through Automotive Safety Integrity Levels (ASIL A–D), where ASIL-D represents the highest safety criticality [4]. For autonomous driving, ISO 21448 (SOTIF) addresses safety of the intended functionality, covering perception failures and unknown unsafe scenarios that fall outside the scope of ISO 26262 [5]. Simultaneously, ISO/SAE 21434 introduces cybersecurity engineering requirements that are now inseparable from functional safety [6]. Historically, safety-critical automotive systems have relied on proprietary architectures, particularly Arm-based lockstep microcontrollers and DSP-centric controllers [7], [8]. RISC-V introduces a fundamentally different paradigm: open ISA governance, transparent extensibility, and vendor independence [9], [10]. This transition is no longer theoretical. In 2025, Andes Technology announced that its D45-SE processor achieved full ISO 26262 ASIL-D certification by SGS-TÜV Saar, integrating dual-core lockstep (DCLS), ECC protection, bus protection, stack protection, and real-time diagnostic safety circuits [11]. Similar developments are visible across the RISC-V automotive ecosystem [12]–[15]. This changes the central research question from “Can RISC-V be used in automotive systems?” to: How can RISC-V become the preferred certification platform for autonomous driving systems? This paper addresses that question by providing a structured analytical framework, a maturity model, and an ML-assisted certification methodology specifically designed for automotive RISC-V platforms.
Scope of This Work. This paper is intended as an analytical perspective paper rather than an experimental implementation study. It does not propose a single isolated algorithmic breakthrough. Instead, it provides a structured analytical framework and expert-driven roadmap for enabling economically superior certification of automotive-grade RISC-V platforms for autonomous driving systems. Its purpose is to structure the research space, identify the highest-value problems, and define a formal basis for certification-oriented RISC-V development. Experimental validation of individual framework components is identified as future work. II. C ONTRIBUTIONS This paper makes the following strategic framework contributions: 1) Certification Economics Framework. We introduce the concept of certification economics for automotive-grade RISC-V systems and formalize it through an analytical cost model, an ML optimization gain model, and a RISC-V Certification Advantage (RCA) score. These are presented as strategic decision-support models for evaluating certification investments, not as experimentally validated predictive models. 2) ML-Assisted Certification Framework. We propose a systematic mapping of ML methods—specifically LLMs with knowledge graphs, reinforcement learning, and graph neural networks—to RISC-V certification bottlenecks. For each primary method, we provide a structured justification of selection over alternatives. ML is treated as an enabler of certification strategy, not as the central research topic. 3) RISC-V Safety Maturity Model (RSMM). We define a five-level maturity assessment framework for evaluating the certifiability of automotive RISC-V platforms, intended to help researchers prioritize work, vendors assess readiness, OEMs evaluate certifiability, and assessment bodies structure expectations. 4) Illustrative Case Study. We apply the framework to an ASIL-D Autonomous Emergency Braking (AEB) ECU scenario, demonstrating where certification cost appears, where RISC-V provides structural advantages, where ML assistance is most impactful, and how the RSMM applies in practice. III. R ELATED W ORK A. Automotive RISC-V Safety The emergence of RISC-V in automotive applications has been documented by several recent efforts. Cuomo et al. [9] present an open RISC-V platform for next-generation automotive ECUs. Andes Technology has achieved both ASIL-D development process certification [12] and ASIL-D product certification for the D45-SE [11] and D23-SE [13] cores. SiFive [14] and Codasip [15] have introduced automotive-grade RISC-V IP with ISO 26262 compliance targets. Pinto et al. [16] identify security gaps in RISC-V MCU architectures, particularly regarding initiator-side protection for mixed-criticality SDV architectures. B. ML for Safety Certification The application of ML to safety engineering is an emerging field. Salay et al. [17], [18] analyze the gaps between ISO 26262 and ML-specific lifecycle requirements. Iyenghar et al. [19] propose systematic enhancements to ISO 26262 with ML-specific testing methods. Cheng et al. [20] provide a quantitative projection of ISO 26262 requirements onto ML-based functions. Kochanthara et al. [21] present safety case patterns for systems with ML components. C. ML Safety in Autonomous Driving Vyas and Xu [1] provide an overview of safety design challenges in AI-driven autonomous vehicles. Werling et al. [2] present a safety integrity framework for automated driving. The SMIRK project [22] demonstrates a complete safety case for an ML component in a pedestrian AEB system. Burton et al. [23] address safety argumentation for ML in highly automated driving. Koopman and Wagner [3] frame autonomous vehicle safety as an interdisciplinary challenge spanning engineering, policy, and validation methodology. D. Formal Verification for RISC-V Reid [24] describes industrial-scale ISA formal verification at Arm, providing a benchmark for RISC-V efforts. Wolf et al. [25] present open-source formal verification frameworks for RISC-V cores. These approaches are directly relevant to ASIL-D certification, where formal correctness proofs can substantially reduce diagnostic coverage requirements [26]. E. Standards Convergence: ISO 26262, ISO 21448, and ISO/SAE 21434 Recent work has highlighted the need to address functional safety, safety of the intended functionality, and cybersecurity as an integrated concern for autonomous driving. Schildbach [27] examines ISO 26262 application in automated vehicle control. Bastos et al. [28] provide a comparative analysis of ISO 21434, ISO 26262, and ML requirements. The SOTIF standard (ISO 21448) [5] addresses perception and decision-making failures that are not covered by ISO 26262’s random hardware fault model, making it essential for autonomous driving systems. The split-and-cover methodology [29] demonstrates SystemC-based FMEDA improvement under ISO 26262. Macher et al. [30] document FMEDA challenges and best practices in automotive contexts.
2
IV. W HY RISC-V I S S TRUCTURALLY D IFFERENT FOR AUTOMOTIVE S AFETY RISC-V should not be analyzed as merely another processor ISA. Its architectural properties create structural advantages for functional safety engineering that are qualitatively different from proprietary alternatives. A. Open ISA and Certification Transparency Unlike proprietary architectures, RISC-V specifications are fully open and auditable [10]. This transparency directly improves traceability from safety requirements to implementation, formal verification feasibility, toolchain qualification transparency, and long-term certification maintainability. In contrast, Arm-based platforms restrict ISA-level access through NDA agreements, limiting independent verification and creating structural opacity in the certification evidence chain [7], [24]. B. Custom Extensions with Controlled Safety Scope Automotive systems increasingly require domain-specific extensions for motor control, cryptographic operations, sensor fusion, deterministic real-time execution, and autonomous driving acceleration. RISC-V enables controlled ISA customization through its standard extension mechanism [10]. However, this introduces a critical certification challenge: customization must not invalidate portability or existing certification evidence. This tension between extensibility and certification stability requires careful architectural governance and represents a uniquely RISC-V research problem. C. Formal Verification Opportunity Because the ISA specification is open, formal verification from ISA semantics to microarchitectural implementation becomes significantly more practical than in closed ecosystems [25]. This is especially valuable for verifying PMP/MMU correctness, privilege isolation, lockstep equivalence, debug access control, and fault containment boundaries—all of which are critical for ASIL-D compliance. D. Toolchain Qualification and Certification Reuse Toolchain qualification represents one of the most significant and least discussed certification cost drivers. Without a qualified compiler, ASIL certification becomes substantially more expensive due to increased verification requirements [4], [31]. RISC-V creates a unique opportunity for qualification-by-construction approaches to compiler certification. The open nature of both the ISA and open-source toolchains (LLVM, GCC) enables: (i) deterministic compilation with traceable optimization passes, (ii) compiler confidence level assessment per ISO 26262 Tool Confidence Level (TCL) requirements, (iii) certification evidence reuse across RISC-V implementations sharing the same ISA subset, and (iv) community-driven qualification artifacts that reduce per-vendor qualification cost [31]. This is in contrast to proprietary architectures, where toolchain qualification is vendor-locked and non-transferable across implementations. E. Security and Functional Safety Convergence Pinto et al. [16] demonstrate that automotive RISC-V systems require stronger security primitives for virtualized MCUs and initiator-side protection, especially for ISO/SAE 21434 compliance in mixed-criticality SDV architectures. This convergence of safety and security is particularly important for autonomous driving platforms where attack surfaces expand with connectivity [32]. F. Structured Comparison: Arm vs. RISC-V Table I presents a systematic comparison across eight functional safety dimensions. TABLE I A RM VS . RISC-V FOR F UNCTIONAL S AFETY C ERTIFICATION Dimension
Arm
RISC-V
ISA Transparency Formal Verification Custom Extensions Toolchain Qualification
Proprietary, NDA-restricted Limited by IP access Vendor-controlled (CIE) Proprietary, vendor-locked
Certification Reuse
Vendor-dependent licensing
Vendor Lock-in Debug Transparency Supply-Chain Independence
High (IP licensing model) Proprietary debug IP Single-vendor dependency
Fully open and auditable Full RTL verification feasible Open standard mechanism Open-source + commercial options Portable across implementations Low (open specification) Open debug specification Multi-vendor ecosystem
3
V. S AFETY R EQUIREMENTS OF AUTONOMOUS D RIVING S YSTEMS Autonomous driving systems introduce fundamentally different safety requirements compared to classical ECUs. The challenge is no longer protecting a single controller but ensuring system-wide safety across the entire autonomous driving stack [1]–[3]. A. Standards Landscape for Autonomous Driving Autonomous driving systems operate at the intersection of three complementary standards: • ISO 26262 addresses random hardware faults and systematic software failures through ASIL classification and diagnostic coverage requirements [4]. • ISO 21448 (SOTIF) addresses safety of the intended functionality—specifically, hazardous behavior caused by functional insufficiencies or foreseeable misuse, including perception uncertainty, unknown unsafe scenarios, sensor ambiguity, planning uncertainty, and ML model uncertainty [5]. • ISO/SAE 21434 addresses cybersecurity engineering, including threat analysis and risk assessment (TARA), which is inseparable from functional safety in connected autonomous vehicles [6]. The convergence of these three standards creates a certification challenge that is qualitatively more complex than traditional ECU certification. RISC-V platforms for autonomous driving must address all three simultaneously. B. Autonomous Driving Safety Stack and RISC-V Implications Table II maps the autonomous driving safety stack to specific RISC-V architectural implications. TABLE II AUTONOMOUS D RIVING S AFETY S TACK AND RISC-V I MPLICATIONS AD Layer
Safety Requirement
RISC-V Implication
Perception
Redundancy
Bounded latency, degraded-safe output under sensor failure Deterministic execution, uncertainty handling ASIL-D integrity, sub-ms response Heterogeneous input (radar, lidar, camera) Fail-operational behavior
Zonal Control
Distributed safety across domains
Degraded Mode
Graceful performance reduction
Deterministic execution, custom sensor fusion extensions Formal verifiability, WCET guarantees DCLS, safety islands, hardware interlocks Multi-modal I/O, DMA protection, PMP isolation Split-lock architecture, independent safety cores Adaptive lockstep, inter-zone isolation Safe-state transition, watchdog supervision
Planning Actuation Sensor Fusion
C. Mixed-Criticality Compute ADAS and autonomous driving combine ASIL-D braking and steering, ASIL-B perception pipelines, QM infotainment workloads, and Linux/RTOS coexistence on shared hardware. This requires strict temporal and spatial isolation with predictable scheduling guarantees [9], [16]. D. Continuous Certification Unlike traditional ECUs, autonomous vehicles are updated continuously via OTA deployment. This creates recurring certification challenges: re-validation, regression detection, compiler re-qualification, and evidence regeneration after each update cycle [19], [28], [33]. VI. P ROPOSED C ERTIFICATION E CONOMICS F RAMEWORK A common misconception is that automotive functional safety is primarily a hardware problem. In practice, the dominant cost drivers are engineering activities: FMEDA/FMEA generation, diagnostic coverage analysis, toolchain qualification, safety case documentation, assessment body audit cycles, evidence traceability, fault injection campaigns, and re-certification after OTA updates [30]. We define certification economics as the optimization of engineering effort, traceability, and compliance cost required to achieve and maintain ASIL certification. The models presented below are intended as analytical decision-support tools for strategic evaluation of certification investments. They provide a structured basis for comparing architectural alternatives and prioritizing automation efforts. Industrial benchmarking and empirical calibration of model parameters are identified as essential future work.
4
A. Certification Cost Function We model the total certification cost as: Ctotal = CFMEDA + Cverif + Ctool + CFI + Caudit + Crecert
(1)
where CFMEDA represents FMEDA and FMEA generation cost, Cverif captures formal and simulation-based verification effort, Ctool denotes toolchain qualification cost, CFI represents fault injection campaign cost, Caudit captures assessment body audit effort, and Crecert represents the recurring cost of re-certification after OTA updates. Figure 1 illustrates the certification economics optimization flow. CFMEDA
Cverif
Ctool
CFI
Caudit
Crecert
Ctotal =
P
Ci
ML-Assisted Optimization: ∆C = Cbaseline − CML-assisted
Optimized Certification Cost + RCA Score Fig. 1. Certification economics optimization flow: cost decomposition, ML-assisted reduction, and advantage scoring.
B. ML Optimization Gain The certification cost reduction achievable through ML-assisted automation is modeled as: ∆C = Cbaseline − CML-assisted
(2)
where Cbaseline is the cost under traditional manual certification workflows and CML-assisted is the cost with ML-augmented processes. The optimization objective is to maximize ∆C while maintaining certification validity and audit confidence. C. RISC-V Certification Advantage Score We define the RISC-V Certification Advantage (RCA) score as a weighted composite metric: RCA = α · T + β · V + γ · F + δ · Q
(3)
where T represents ISA transparency (auditable specification access), V denotes formal verifiability (feasibility of ISA-toRTL proofs), F captures formal proof feasibility (coverage of safety properties), and Q quantifies qualification cost reduction (toolchain and evidence reuse). The coefficients α, β, γ, δ ∈ [0, 1] with α + β + γ + δ = 1 are context-dependent and determined by the specific ASIL target and certification scope. The RCA score is a structured evaluation tool, not a predictive formula. Its value lies in enabling systematic comparison of architectural alternatives for certification cost optimization. VII. ML-A SSISTED C ERTIFICATION F RAMEWORK Machine Learning is not the central topic of this paper. Rather, ML is treated as an enabling technology that addresses specific bottlenecks in RISC-V certification workflows. The guiding question is: How can ML reduce the cost and uncertainty of certifying automotive-grade RISC-V systems? We focus on three primary ML integration points where the strongest certification impact is expected, with structured justification for each.
5
A. LLM + Knowledge Graph for FMEDA and Safety Case Automation Problem. FMEDA generation remains largely manual, spreadsheet-driven, and expensive [29], [30]. Safety documentation requires maintaining traceable links from hazard analysis (HARA) through FMEA, FMEDA, verification evidence, and compliance mapping to specific ISO 26262 clauses. Why LLMs + Knowledge Graphs. Large Language Models can process heterogeneous technical documents (RTL specifications, safety manuals, requirements databases) and generate structured outputs including preliminary FMEDA drafts, failure mode extractions, and requirement-to-failure traceability mappings [34]. However, LLMs alone suffer from hallucination risk that is unacceptable in safety-critical certification contexts. We therefore propose a combined architecture: a certification knowledge graph provides the structural backbone encoding the traceability chain Requirement → Implementation → V erif ication → Evidence → ISO Clause, while LLM-based agents perform document parsing and draft generation with graph-grounded retrieval to mitigate hallucination [21], [23]. Why not alternatives. Rule-based systems lack flexibility across diverse RISC-V implementations. Traditional NLP approaches (BERT, Word2Vec) support specific subtasks [34] but lack generative capability. LLMs without knowledge graph grounding introduce unacceptable hallucination risk for certification evidence. B. Reinforcement Learning for Fault Injection Optimization Problem. ISO 26262 requires extensive fault injection testing. Brute-force campaigns across the full fault space are prohibitively expensive for complex automotive SoCs [4], [30]. Why RL. Fault injection optimization is a sequential decision problem: at each step, the agent selects the next fault to inject based on observed coverage improvement. This maps naturally to a Markov Decision Process where the state represents current diagnostic coverage, actions are fault selections, and the reward is coverage gain per injection cost. Why not supervised learning. There are no fixed ground-truth labels for optimal injection sequences. Why not static heuristics. Static strategies provide poor exploration of complex fault spaces and cannot adapt to implementation-specific coverage distributions. C. GNNs for Diagnostic Coverage Optimization Problem. Diagnostic coverage is often the deciding factor for ASIL compliance. Optimizing safety mechanism placement requires understanding fault propagation across hierarchical hardware structures [29]. Why GNNs. The hardware architecture of an automotive SoC is naturally represented as a graph. Graph Neural Networks can model fault propagation paths, predict latent fault manifestation, optimize safety monitor placement, and determine optimal supervision boundaries for lockstep systems, safety islands, and mixed-criticality SoCs. Why not traditional analysis. Manual fault propagation analysis does not scale to modern SoC complexity. Statistical methods lack the structural awareness that graph-based approaches provide. D. ML Method Summary Table III summarizes the primary and secondary ML integration points. TABLE III ML-A SSISTED C ERTIFICATION : M ETHOD M APPING AND P RIORITY Certification Problem
ML Method
Priority
FMEDA generation Safety case automation Fault injection optimization Diagnostic coverage analysis Compiler qualification
LLM + NLP Knowledge Graph + LLM Reinforcement Learning Graph Neural Networks Anomaly Detection
Very High Very High Very High Very High High
E. Trustworthiness and Limitations of ML in Certification The application of ML to safety certification introduces inherent trustworthiness challenges that must be explicitly acknowledged: • LLM hallucination risk. Generated FMEDA entries or safety case fragments may contain plausible but factually incorrect content. All ML-generated certification artifacts require expert review and formal validation before inclusion in audit evidence. • Certification authority acceptance. Assessment bodies (e.g., TÜV, UL) have not yet established formal acceptance criteria for ML-generated certification evidence. Regulatory framework development is a prerequisite for production deployment of ML-assisted certification [35]. • Validation of ML-generated evidence. ML outputs used in certification must themselves be validated to a confidence level commensurate with the target ASIL. This creates a meta-certification challenge.
6
OEM trust constraints. Automotive OEMs require deterministic, auditable processes. ML-assisted workflows must provide full traceability and explainability to meet OEM governance requirements. • Proprietary data limitations. Training ML models for certification tasks requires access to FMEDA databases, safety cases, and fault injection results that are typically proprietary and commercially sensitive. These limitations do not invalidate the proposed framework but define the boundaries within which ML-assisted certification can be responsibly deployed. •
VIII. RISC-V S AFETY M ATURITY M ODEL To provide a structured evaluation framework for automotive RISC-V platforms, we propose a five-level RISC-V Safety Maturity Model (RSMM). The RSMM is designed as a maturity assessment tool rather than a strict quantitative instrument. Its purpose is to enable: • Researchers: prioritize work toward the highest-impact certification gaps. • Vendors: assess readiness of their RISC-V IP for specific ASIL targets. • OEMs: evaluate certifiability of candidate platforms for vehicle programs. • Assessment bodies: structure certification expectations for RISC-V-based systems. Table IV defines the five levels. TABLE IV RISC-V S AFETY M ATURITY M ODEL (RSMM) Level
Classification
Characteristics
1
Safety-Aware MCU
2
ASIL-B Capable System
3
ASIL-D Lockstep Platform
4
Mixed-Criticality AD Platform
5
Continuous-Certification OTA Platform
Basic safety features (ECC, watchdog), no formal ASIL target, suitable for QM applications. Hardware fault metrics meet ASIL-B, partial FMEDA, basic diagnostic coverage, initial toolchain qualification. Full DCLS, comprehensive FMEDA, diagnostic coverage ≥ 99%, qualified toolchain, TÜV-assessed safety manual. Hypervisor-certified isolation, ASIL-D + ASIL-B + QM coexistence, safety islands, deterministic scheduling, SOTIF-aware design. Incremental re-certification, ML-assisted evidence regeneration, regression-aware validation, live safety case updates.
Figure 2 illustrates the RSMM progression. Level 1 Safety-Aware MCU
Level 2: ASIL-B Capable
Current: Andes D45-SE
Level 3: ASIL-D Lockstep Platform
Level 4: Mixed-Criticality AD Platform
Level 5: Continuous-Certification OTA Platform Fig. 2. RISC-V Safety Maturity Model progression. Current commercially certified platforms correspond to Level 3.
Current commercially certified platforms such as the Andes D45-SE correspond to Level 3 [11]. The transition from Level 3 to Level 5 represents the primary research and engineering challenge for the RISC-V automotive ecosystem. IX. RISC-V S AFETY A RCHITECTURE PATTERNS Three dominant architectural patterns emerge for ASIL-D-capable RISC-V platforms.
7
A. Dual-Core Lockstep (DCLS) DCLS remains the dominant ASIL-D safety mechanism, in which two identical cores execute the same instruction stream while comparator logic detects divergence. The Andes D45-SE integrates DCLS with real-time diagnostic circuits [11]. Open research directions include delayed lockstep, split-lock architectures [13], heterogeneous lockstep, and adaptive lockstep for zonal controllers. B. Safety Islands Dedicated safety cores supervise watchdog functions, emergency state transitions, fault detection, and safety policy enforcement. Safety islands are critical for centralized autonomous platforms where independent supervision of the main compute cluster is required [8]. C. Mixed-Criticality Isolation Automotive RISC-V platforms require hypervisor-certified separation for Linux and AUTOSAR coexistence, deterministic interrupt behavior, and memory protection through PMP and hardware partitioning [16]. This represents a system-level challenge rather than a core-level problem. X. I LLUSTRATIVE C ASE S TUDY: ASIL-D AEB ECU To demonstrate practical application of the proposed framework, we present an illustrative scenario based on an ASIL-D Autonomous Emergency Braking (AEB) ECU. This is not an experimental validation but an application example that grounds the analytical framework in a realistic certification context. A. System Description Consider an AEB ECU based on a RISC-V DCLS platform (RSMM Level 3) responsible for: (i) radar and camera sensor fusion, (ii) collision risk assessment, (iii) emergency braking actuation at ASIL-D integrity, and (iv) fail-safe degradation under sensor failure. B. Certification Cost Analysis Applying the cost model (Eq. 1): • CFMEDA : The AEB ECU requires comprehensive FMEDA covering the DCLS core, comparator logic, ECC memory, bus protection, and safety island. This is the highest single cost component, typically requiring 6–12 months of expert engineering effort. • Ctool : Compiler qualification for the RISC-V toolchain (GCC or LLVM) at TCL-2 or TCL-3 is required. RISC-V’s open toolchain enables qualification evidence reuse, reducing this cost compared to proprietary alternatives (Section IV-D). • CFI : Fault injection campaigns must cover the full DCLS fault space. RL-based optimization (Section VII) targets this cost component. • Caudit : TÜV assessment requires complete traceability from requirements to evidence. Knowledge-graph-based automation directly reduces audit preparation cost. • Crecert : OTA firmware updates to the AEB ECU trigger re-certification requirements, making this a recurring cost that grows with vehicle lifetime. C. RISC-V Advantage Assessment For this AEB scenario, the RCA score (Eq. 3) is dominated by: High T (transparency): open ISA enables independent FMEDA verification. • High V (verifiability): DCLS equivalence is formally provable. • High Q (qualification reuse): compiler qualification evidence is portable across RISC-V implementations. •
D. RSMM Application The AEB ECU requires RSMM Level 3 (ASIL-D Lockstep Platform) as a minimum. If the ECU must support OTAupdateable perception models, it approaches Level 4 (Mixed-Criticality AD Platform) requirements. The RSMM provides a structured basis for evaluating whether a candidate RISC-V platform meets the certification readiness requirements for this application. XI. E VALUATION M ETHODOLOGY We define the evaluation metrics that would be used to assess each component of the framework in future industrial deployment.
8
A. Certification Cost Metrics FMEDA effort reduction (%): Reduction in person-hours for FMEDA generation using LLM-assisted automation compared to manual baseline. • Certification cycle time reduction: Reduction in calendar time from project start to TÜV assessment completion. • Fault injection efficiency: Diagnostic coverage achieved per unit of fault injection campaign cost (coverage-per-dollar). • Audit effort reduction: Reduction in assessor review hours attributable to improved evidence traceability. • Re-certification latency: Time from OTA update deployment to updated certification evidence availability. •
B. ML Component Metrics Traceability completeness: Percentage of requirements with verified bidirectional links to implementation, verification, and evidence. • False positive rate: Rate of incorrect faults or spurious safety mechanism recommendations by ML components. • Safety case coverage: Percentage of ISO 26262 clauses with automatically generated evidence linkage. •
XII. P ROPOSED C ERTIFICATION W ORKFLOW Figure 3 presents the proposed seven-layer certification workflow for automotive RISC-V platforms. Vehicle Layer: Autonomous Driving Platform
RISC-V Platform Layer: DCLS + Safety Island + Isolation
Safety Architecture: ISO 26262 + SOTIF + ISO/SAE 21434
Certification: FMEDA + Fault Injection + Verification
ML Assistance: LLM + KG + RL + GNN
Evidence: Traceability + Safety Case + Audit Package
Regulatory Approval: TÜV Assessment + ASIL-D Certification Fig. 3. Seven-layer certification workflow for automotive RISC-V autonomous driving platforms, integrating ISO 26262, ISO 21448 (SOTIF), and ISO/SAE 21434.
XIII. D ISCUSSION AND L IMITATIONS A. Research Implications Traditional academic architecture research focuses on IPC, cache optimization, branch prediction, and microarchitectural novelty. For automotive RISC-V, these considerations are secondary. The highest-value research contribution lies in certification infrastructure rather than processor performance. The commercially successful RISC-V automotive platform may not be the fastest. It may be the platform that is easiest to certify—the one with the strongest compiler qualification package, the most comprehensive TÜV evidence package, the clearest safety manual, and the fastest path to ASIL-D production deployment. This represents a fundamental shift in how automotive processor research should be funded, evaluated, and commercialized.
9
B. Positioning as an Analytical Perspective This work does not propose a single isolated algorithmic breakthrough. Instead, it provides a structured analytical framework and expert-driven roadmap for enabling economically superior certification of automotive-grade RISC-V platforms. The certification economics model, RSMM, and ML-assisted certification framework are intended as strategic evaluation tools that structure the research space and identify high-value problems. This positioning is appropriate for an interdisciplinary domain characterized by emerging standards, industrial transformation, and early-stage ecosystem development. C. Limitations The analytical framework has not been empirically validated through industrial case studies. The cost model parameters require calibration against real certification projects. • The RSMM levels are defined based on expert analysis of current certification practice. As the RISC-V automotive ecosystem matures, the model may require refinement. • ML-assisted certification faces significant barriers to industrial adoption, including assessment body acceptance, hallucination risk, and the meta-certification challenge of validating ML-generated evidence (Section VII-E). • The illustrative case study (Section X) demonstrates framework applicability but does not constitute experimental validation. •
D. Future Work Priority future work includes: (i) industrial pilot deployment of the ML-assisted FMEDA generation pipeline on a specific RISC-V implementation, (ii) empirical calibration of the certification cost model using data from real ASIL-D projects, (iii) development of formal acceptance criteria for ML-generated certification evidence in collaboration with assessment bodies, and (iv) extension of the RSMM to incorporate SOTIF and cybersecurity maturity dimensions. XIV. C ONCLUSION RISC-V automotive adoption has entered a phase where ASIL-D certification demonstrates technical feasibility [11]. The strategic challenge is no longer building safer processors but making safety certification economically superior—especially for autonomous driving systems that must simultaneously address ISO 26262, ISO 21448 (SOTIF), and ISO/SAE 21434. This paper contributes a structured analytical framework comprising certification economics models, a five-level RISC-V Safety Maturity Model, and an ML-assisted certification methodology that maps LLMs with knowledge graphs, reinforcement learning, and graph neural networks to specific certification bottlenecks. The structured Arm vs. RISC-V comparison demonstrates that RISC-V provides structural advantages in ISA transparency, formal verifiability, toolchain qualification portability, and certification evidence reuse. The future of automotive functional safety is not safer CPUs—it is certifiable platforms. The future of certification is not manual compliance work—it is structured, traceable, and intelligently assisted certification infrastructure. RISC-V is uniquely positioned to become that platform. The strongest outcome of this research direction is not an academic paper but a deployable platform: a certified, secure, traceable, ASIL-D-ready RISC-V autonomous driving platform. This is where both academic impact and industrial value converge. R EFERENCES [1] V. Vyas and Z. Xu, “Key Safety Design Overview in AI-Driven Autonomous Vehicles,” arXiv preprint, vol. arXiv:2412.08862, 2024. [2] M. Werling, J. Betz, and M. Lienkamp, “Safety Integrity Framework for Automated Driving,” arXiv preprint, vol. arXiv:2503.20544, 2025. [3] P. Koopman and M. Wagner, “Autonomous Vehicle Safety: An Interdisciplinary Challenge,” IEEE Intelligent Transportation Systems Magazine, vol. 9, no. 1, pp. 90–96, 2019. [4] ISO 26262-2018: Road Vehicles—Functional Safety, International Organization for Standardization Std., 2018. [5] ISO 21448:2022: Road Vehicles—Safety of the Intended Functionality (SOTIF), International Organization for Standardization Std., 2022. [6] ISO/SAE 21434:2021: Road Vehicles—Cybersecurity Engineering, International Organization for Standardization Std., 2021. [7] Arm Holdings, Arm Cortex-R52+ Processor for ASIL-D Safety-Critical Applications, Technical Reference Manual, 2023. [8] Infineon Technologies, “AURIX TC4x: Next-Generation Automotive Microcontroller for ADAS and Domain Control,” Product Brief, 2023. [9] L. Cuomo, F. Casini, and M. Luise, “Towards a RISC-V Open Platform for Next-Generation Automotive ECUs,” arXiv preprint, vol. arXiv:2307.04148, 2023. [10] A. Waterman and K. Asanović, The RISC-V Instruction Set Manual, Volume I: User-Level ISA, Version 2.2, RISC-V Foundation, 2019. [11] Andes Technology, “D45-SE Processor Achieves ISO 26262 ASIL-D Certification by SGS-TÜV Saar,” Press Release, 2025. [12] ——, “First RISC-V Vendor to Accomplish ISO 26262 ASIL-D Development Process Certification,” 2022. [13] ——, “D23-SE Functional Safety Core with DCLS and Split-Lock Architecture,” Product Brief, 2025. [14] SiFive, “Automotive-Grade RISC-V IP for ISO 26262 Functional Safety Applications,” 2023. [15] Codasip, “Customizable RISC-V Processors for Safety-Critical Automotive Applications,” Product Brief, 2024. [16] S. Pinto, H. Santos, and D. Oliveira, “RISC-V Needs Secure Wheels: The MCU Initiator-Side Perspective,” arXiv preprint, vol. arXiv:2410.09839, 2024. [17] R. Salay, R. Queiroz, and K. Czarnecki, “An Analysis of ISO 26262: Using Machine Learning Safely in Automotive Software,” arXiv preprint, vol. arXiv:1709.02435, 2017. [18] R. Salay and K. Czarnecki, “Using Machine Learning Safely in Automotive Software: An Assessment and Adaption of Software Process Requirements in ISO 26262,” arXiv preprint, vol. arXiv:1808.01614, 2018.
10
[19] P. Iyenghar et al., “A Systematic Approach to Enhancing ISO 26262 with Machine Learning-Specific Life Cycle Phases and Testing Methods,” IEEE Access, vol. 12, 2024. [20] C. H. Cheng et al., “Quantitative Projection of ISO 26262 on ML-Based Functions in Automated Driving,” in Proc. SafeAI Workshop. AAAI, 2020. [21] S. Kochanthara et al., “A Safety Case Pattern for Systems with Machine Learning Components,” in Proc. Ada-Europe, 2021, pp. 129–143. [22] M. Borg et al., “Ergo, SMIRK Is Safe: A Safety Case for a Machine Learning Component in a Pedestrian Automatic Emergency Brake System,” Software Quality Journal, vol. 30, pp. 335–367, 2022. [23] S. Burton, L. Gauerhof, and C. Heinzemann, “Making the Case for Safety of Machine Learning in Highly Automated Driving,” in Proc. SAFECOMP Workshops, 2017, pp. 5–16. [24] A. Reid, “Trustworthy Specifications of ARM v8-A and v8-M System Level Architecture,” in Proc. FMCAD, 2016, pp. 161–168. [25] C. Wolf et al., “A RISC-V Formal Verification Framework,” in Proc. RISC-V Workshop, 2018. [26] IEC 61508:2010: Functional Safety of Electrical/Electronic/Programmable Electronic Safety-Related Systems, International Electrotechnical Commission Std., 2010. [27] G. Schildbach, “On the Application of ISO 26262 in Control Design for Automated Vehicles,” arXiv preprint, vol. arXiv:1804.04349, 2018. [28] S. Bastos, K. Castelo Branco, and A. Oliveira, “Bridging the Gaps: A Comparative Analysis of ISO 21434, ISO 26262 and Machine Learning in Autonomous Vehicles,” in Proc. Brazilian Symp. Robotics, 2025. [29] A. Becker et al., “Split’n’Cover: ISO 26262 Hardware Safety Analysis with SystemC,” Int. J. Parallel Programming, vol. 53, 2025. [30] G. Macher et al., “Automotive FMEDA: Challenges and Best Practices,” in Proc. SAFECOMP, 2019, pp. 299–314. [31] Andes Technology, “HighTec C/C++ Compiler Suite Supports Andes ISO 26262 Certified RISC-V IP,” Press Release, 2024. [32] R. Bosch GmbH, “Safety and Security for Automated Driving: Challenges and Solutions,” in Proc. SAE World Congress, 2021. [33] NHTSA, “Framework for Automated Driving System Safety,” U.S. Department of Transportation, 2022. [34] M. Kim et al., “NLP-Based Fault Test Case Generation for ISO 26262 Validation Using BERT and Word2Vec,” Sensors, vol. 24, no. 3, 2024. [35] UL Solutions, “Functional Safety Assessment for RISC-V Processor IP,” Assessment Report, 2023.
11