
Copyright and License: © The Author(s). The articles are published under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0), which permits sharing, adaptation, and reuse with appropriate credit to the original authors and source. Any adapted or derivative work must be distributed under the same CC BY-SA 4.0 license.
Knowledge-Grounded Reasoning, Trajectory Refinement, and Hardware-Efficient Deployment in Autonomous Driving: A Layered Survey Across Three Generations of AI
By: Gaurav Pokharkar | Atharva Khandve | Gopi Gaddamadugu | Akshay Choudhari
| Pages: 18 - 36
|
Open
Abstract
Autonomous driving has evolved through three overlapping generations of artificial intelligence: rule-based controllers, deep-learning perception and policy networks, and knowledge-grounded systems built on large language models (LLMs), vision-language models (VLMs), and vision-language-action (VLA) models. This review covers 107 records (90 autonomous-driving contributions and 17 adjacent foundational sources), retained after a citation-integrity re-audit of nine database and proceedings searches spanning January 2016 through June 2026. We organize the evidence around three interacting pillars: knowledge-grounded reasoning; trajectory generation and refinement; and hardware-aware execution. A two-axis taxonomy classifies methods by functional role and by an operational runtime class (hard real-time, soft real-time, cognitive, or not demonstrated). Foundation-model methods primarily support semantic interpretation, memory, explanation, and high-level planning; dynamically feasible motion still depends on constrained planning, checking, and fallback mechanisms. Most LLM- and VLM-based studies report neither automotive-platform latency nor worst-case timing, and heterogeneous hardware and evaluation protocols across the reported benchmarks preclude a defensible global ranking. We therefore propose a layered design principle: cognitive outputs should enter the driving stack as checked goals or constraints, while requirement-derived deadlines and runtime assurance remain the responsibility of faster planning, monitoring, and control layers.
DOI URL: https://doi.org/10.64820/AEPJMLDL.32.18.36.122026





