A multicriteria decision making approach to study barriers to the adoption of autonomous vehicles

The paper addresses uncertainty around the large-scale adoption of autonomous vehicles (AVs) despite their potential benefits (safety, congestion reduction, fuel savings). It highlights the need to understand and prioritize barriers—both physical (infrastructure) and psychological (public perception)—that impede AV diffusion, especially during a transition period where AVs and conventional vehicles coexist. The study critiques prior work for focusing on pairwise, largely associational relationships between barriers and proposes to model causal, system-level interdependencies. It poses three research questions: (a) What are the key barriers to AV adoption? (b) How do these barriers causally influence each other? (c) How can these causal influences be depicted and analyzed? To answer these, it combines Grey-DEMATEL with systems thinking (causal loop diagrams) using expert judgments within the US context, where AV readiness slipped in 2019, indicating salient barriers.

The literature review frames AV research growth post-2014 across themes: opportunities and challenges for transport policy; consumer willingness-to-pay, behavior, and risk perceptions; and system-level impacts (parking design, fuel consumption). Methodologically, most studies rely on stated preference surveys with econometric analyses and simulations to forecast adoption and impacts (e.g., fuel use, travel behavior). Expected benefits reported include reduced crashes, emissions, congestion, driving stress, and improved mobility for vulnerable groups; however, some benefits are debated due to induced travel. A synthesis table summarizes opportunities and barriers identified by prior studies, including security/privacy, legal liability, standards and certification, infrastructure needs, and willingness to pay. The review identifies a gap: lack of comprehensive causal analysis among multiple barriers. Using a three-step process—Scopus search (87 hits; 42 reviewed), synthesis of distinct barriers (10), and expert validation (6 experts)—the study defines ten barriers: reduced security/privacy (RSP), social inequity (SIN), obscurity in accountability (OSA), lack of customer acceptance (LCA), potential loss of employment (PLE), inadequate infrastructure (INF), lack of standards (LOS), absence of regulation and certification (ARC), manufacturing cost (MNC), and induced travel (ITRL).

The research employs a five-stage approach: (1) identify key barriers via literature and expert validation; (2) collect pairwise influence assessments from experts; (3) apply Grey-DEMATEL to quantify causal structure, rank barriers (prominence R+C) and classify as cause/effect (net influence R−C); (4) conduct sensitivity analyses with alternative expert weightings; (5) depict and interpret causal influences via a causal loop diagram (CLD). The US context was selected for its leading AV innovation but slipping readiness. Expert elicitation: 55 experts contacted; 18 completed responses (14 academia, 4 industry), most with >3.5 years AV experience. Experts rated the influence of each barrier on others using a six-level linguistic scale mapped to grey numbers: N [0,0], VL [0,1], L [1,2], M [2,3], H [3,4], VH [4,5], yielding 18 direct-relation 10×10 matrices. Grey-DEMATEL procedures (per Bai & Sarkis and others) included: converting linguistic inputs to grey matrices; averaging to an aggregate grey-relation matrix; normalization; defuzzification to crisp values; forming a normalized direct relation matrix X; computing the total relation matrix M = X(I−X)−1; deriving R (row sums) and C (column sums); calculating R+C (prominence) and R−C (net influence) for classification; and plotting an Influence-Prominence Map. For CLD construction, only influences above a threshold θ were drawn to avoid clutter; θ was set to μ+σ of M’s off-diagonal entries, computed as 0.0665, yielding 17 above-threshold influences forming 10 feedback loops. Sensitivity analysis reweighted experts by experience group across six scenarios (e.g., 50/30/20% for >8, 5–8, 3.5–5 years) to test robustness of ranks (R+C and R−C) and above-threshold links.

- Prominence (R+C): Lack of customer acceptance (LCA) is most prominent (R+C = 1.175; rank 1). Next are lack of standards (LOS, 0.900; rank 2) and absence of regulation and certification (ARC, 0.850; rank 3). Obscurity in accountability (OSA, 0.714; rank 4), inadequate infrastructure (INF, 0.708; rank 5), manufacturing cost (MNC, 0.695; rank 6), reduced security and privacy (RSP, 0.590; rank 7), induced travel (ITRL, 0.541; rank 8), social inequity (SIN, 0.494; rank 9), and potential loss of employment (PLE, 0.132; rank 10). - Net influence (R−C) classification: Cause (positive R−C): MNC (0.134; rank 1), LOS (0.102; rank 2), OSA (0.093; rank 3), INF (0.088; rank 4), RSP (0.040; rank 5), PLE (0.005; rank 6). Effect (negative R−C): ARC (−0.012; rank 7), SIN (−0.068; rank 8), ITRL (−0.173; rank 9), LCA (−0.209; rank 10). - CLD structure: With threshold θ = 0.0665, 17 above-threshold causal links formed 10 feedback loops. LCA is central, participating in 9 causal relationships and 6 of the 10 loops, indicating its systemic centrality. Notable mutual influences include LCA ↔ RSP (privacy/security) and LCA ↔ ARC (regulation/certification). ARC also drives OSA and LOS, which in turn influence INF and LCA, creating multi-step feedback (e.g., LCA → ARC → LOS → INF → LCA; LCA → ARC → LOS → OSA → LCA). - Policy-relevant levers: Despite LCA being the most prominent barrier, more tangible upstream levers—LOS and ARC—are influential drivers affecting LCA and other barriers. MNC exerts the strongest net outward influence, directly affecting LCA and SIN. - Sensitivity and robustness: Across six alternative expert-weighting scenarios, R+C ranks changed minimally (maximum shift: 3 positions for ITRL; LCA, PLE, MNC unchanged). R−C ranks also showed small variability (MNC top and LCA bottom unchanged). The number of above-threshold relationships varied modestly (17–20), with 16 of the 17 base links appearing in at least four scenarios, supporting robustness of the causal structure.

The findings address the research questions by: (a) identifying ten salient AV adoption barriers; (b) quantifying and classifying their causal interdependencies using Grey-DEMATEL; and (c) visualizing the dynamic structure via a CLD. LCA emerges as the most prominent, strongly interconnected barrier; however, the system perspective reveals that improving LOS (standards) and ARC (regulation/certification)—upstream cause factors—can indirectly and directly enhance LCA, reduce OSA, and strengthen INF. MNC’s strong outward influence suggests cost reductions can improve LCA and mitigate SIN. The feedback loops illustrate how regulatory clarity promotes standards, reduces accountability obscurity, facilitates infrastructure development, and builds public trust, creating reinforcing dynamics for AV adoption. Thus, policymakers and industry should prioritize establishing consistent regulations and standards, alongside cost-reduction strategies and privacy/security assurances, to catalyze broader acceptance and deployment.

The study contributes by: (1) systematically identifying and ranking AV adoption barriers; (2) revealing causal, not merely associative, interdependencies among them; and (3) integrating Grey-DEMATEL with systems thinking to map feedback loops. Key insights are that LCA is the most prominent barrier, but tangible levers—industry standards and regulatory/certification frameworks—should be prioritized to indirectly improve acceptance. Cost reduction can further alleviate acceptance and equity concerns. Practical implications include coordinated actions by government (uniform regulations, testing/certification standards), industry (standardization, transparency, cybersecurity, cost reduction), and potential deployment of shared AV models. Future research should extend the expert base and geographic scope, and integrate DEMATEL with econometric adoption models to triangulate consumer-level evidence with system-level causality.

- Expert-based judgments from 18 US-focused experts may limit generalizability; larger and more diverse panels could reveal additional nuances. - Equal-weight base scenario may not fully capture expertise heterogeneity, though sensitivity analyses suggest robustness. - Potentially important barrier of community acceptance (beyond individual consumer acceptance) was not included but may influence adoption. - As an expert-elicitation and systems modeling study, empirical validation with observed adoption data is outside scope; future mixed-method approaches are suggested.