TinyML-Enabled Internet of Vehicular Things: A Cross-Layer Survey of Architectures, V2X Co-Design, Trustworthiness, Benchmarking, and Future Directions

Tiny Machine Learning (TinyML) enables resource-constrained devices to perform local inference in the Internet of Vehicular Things (IoVT), potentially reducing communication overhead, latency, and dependence on cloud services. This survey systematically reviews 57 primary studies covering vehicular applications, embedded hardware, lightweight learning models, optimization techniques, and deployment practices. The reviewed studies are organized using a seven-dimensional literature taxonomy covering deployment location, data modality, learning task, learning paradigm, optimization strategy, application domain, and system objective. Separately, their evaluation practices are critically examined through predictive performance, memory usage, inference latency, energy consumption, communication requirements, trustworthiness, and deployment evidence. The taxonomy characterizes existing research, whereas the benchmarking dimensions support comparison of reported experimental results and deployment limitations. The survey also examines computation–communication trade-offs, V2X interaction, collaborative learning, security, privacy, robustness, and functional safety. The findings reveal substantial heterogeneity in datasets, hardware platforms, evaluation protocols, and resource measurements, limiting direct cross-study comparisons. Moreover, experimentally demonstrated embedded feasibility does not necessarily establish reliability under realistic vehicular conditions or compliance with functional-safety requirements. The survey concludes by identifying research gaps in standardized benchmarking, hardware-aware optimization, adaptive learning, real-world validation, and trustworthy distributed vehicular intelligence.