Explainable Artificial Intelligence for Autonomous Driving: A Comprehensive Overview and Field Guide for Future Research Directions

In plain words

Self-driving cars make split-second decisions, and today's AI usually cannot say why it chose a given action. That matters for safety, for public trust, and for the laws that will govern these vehicles. This paper is a field guide to the ways researchers have tried to make such decisions explainable. It groups the work by technique: highlighting the image regions a model looked at, training agents that can justify themselves, using simple decision trees or logic rules, asking people what explanations they want, and using large language and vision-language models that talk about the scene.

The authors also point out a gap: most earlier reviews ignored when an explanation should be delivered. Their view is that explanations should arrive before a car acts, not after. Their proposal is a conceptual design that combines a driving controller, a regulatory check and an explanation layer, together with a list of open problems for the next generation of vehicles. It is a survey and a position piece, not a system you can run.

TL;DR

• A survey of explainable AI (XAI) for autonomous driving, with six method families in tables tab. 1 tab. 2 tab. 3 tab. 4 tab. 5 tab. 6.
• Argues that the timing of explanations, not only their content, has been neglected; cites takeover studies to suggest a lead time of a few seconds §5.
• Proposes a conceptual framework that combines an end-to-end control component, a safety and regulatory component and an explanation component §5 fig. 7.
• Contains no experiments. Use it to find the literature and open problems, not for numbers to reproduce.

Problem

Autonomous vehicles act on AI decisions that people cannot readily inspect, which the authors say undermines social acceptance and regulatory compliance §1. They cite the GDPR "right of explanation", the EU trustworthy-AI requirements, and the ISO 26262 and ISO 21448 standards as pressure for explainability §2.3. They note that modular pipelines are easier to explain and debug, but errors accumulate across modules, whereas end-to-end driving unifies perception, localization, planning and control in one learned task §2.1. Existing reviews cover form and content of explanations but not their timing, the rise of transformers, LLMs and vision-language models, or a method-based classification §4.1.

Method

The paper has three parts. Each is summarised below.

A. Survey of methods §4. Studies are grouped by family and placed in one table per family:

1. Visual explanations (CNN saliency, attention, semantic segmentation, counterfactual image edits) §4.3 tab. 1 fig. 4 fig. 5.
2. Reinforcement and imitation learning with interpretable components (semantic prediction, bird's-eye-view masks, decision-tree policies, state representations) §4.4 tab. 2 fig. 6.
3. Decision-tree explanations, including goal recognition and random forests with SHAP values §4.5 tab. 3.
4. Logic-based explanations, including signal temporal logic and answer set programming §4.6 tab. 4.
5. User-study-based explanations on what drivers and road users need §4.7 tab. 5.
6. LLM- and VLM-based explanations, such as live natural-language commentary, video question answering and chain-of-thought control §4.8 tab. 6.

B. Working definition §4.8. The paper defines explainable autonomous driving as a self-driving approach that (1) keeps the real-time decisions at an acceptable safety level, (2) gives explanatory information on action decisions in critical scenarios in a timely way, and (3) obeys the traffic rules set by legal authorities. This sentence is quoted in the paper as a definition, so treat its wording as the authors'.

C. Conceptual framework §5 fig. 7. Three components:

• An end-to-end control component. A control system maps the set of environment instances E to the set of actions A. The paper calls it end-to-end (eeC) when this map is total, meaning every environment instance gets an action §5. The two learning paradigms named are reinforcement learning and imitation learning (behaviour cloning and inverse RL) §5.
• A safety and regulatory compliance component (srC). It is written as a function f of the controller and the action set, srC = f(eeC, A), and represents an agency that verifies safety §5. Two routes are named: compliance by simulation and compliance by verification on real roads §5. ISO 26262 and ISO/SAE 21434 are the standards the srC must meet §5.
• An explanation component. It justifies the action choices, can be shown in visual, textual, feature-importance or multimodal form, and is delivered during the trip or in post-trip analysis §5.

The framework also analyses timing in three questions: whether explanations come before or after an action, how much lead time a takeover needs, and whether explanations should be continuous or only when needed §5. For these the paper gives the conclusions in the Hyperparameters section below.

D. Challenges for "AV 2.0" §6 fig. 8. The paper lists safety problems from Amodei et al.'s five AI-safety issues (side effects, reward hacking, scalable oversight, safe exploration, distributional shift) applied to end-to-end driving, plus the missing "fail-safe" ability, and an explainability gap around timing, hallucination and adversarial questions §6.

The interfaces of the framework, written as a sketch. This is not an algorithm from the paper, only a restatement of the components in §5:

# Interfaces of the conceptual framework in [§5]. Not an algorithm in the
# paper; the paper defines these as sets and maps.
# E: set of environment instances; A: set of actions        [§5]

def eeC(environment):
    # Total map: every environment instance gets one action  [§5]
    # The paper gives no construction for this map.
    return action  # action in A

def srC(eeC, action_set):
    # Compliance function f(eeC, A), checked before deployment [§5]
    # f is not specified in the paper.
    return compliant  # True or False

def explain(environment, action, timing):
    # Explanation for the chosen action; timing: "before" or "after"  [§5]
    # Paper prefers "before the action is decided" [§5]
    return text_or_visual_explanation

Hyperparameters

The paper trains nothing, so there are no training hyperparameters. The rows below are the design numbers it reports from cited user studies.

Results

Author-reported. The paper runs no experiments of its own. The table gives headline numbers it quotes from cited studies, with the setup it states. Check each original before you rely on it.

Paper vs code

No official code. The paper gives no code link, and it is a survey without an implementation, so there is no repository to compare. We did not run a GitHub search for this card. Survey repositories found on GitHub are typically paper lists, not implementations, and are not used here.

Pitfalls

• Nothing to reproduce. All numbers come from other studies and are given as the authors describe them. Their datasets, metrics and setups differ. Check each original before you rely on a number.
• The framework is informal. The equations in §5 are not numbered in the paper, so cite them by section. The total map eeC is required to cover every environment instance, while §6 itself treats distributional shift (transfer from simulation to real roads) as an open problem. The paper does not reconcile the two.
• f in srC is undefined. The paper says srC = f(eeC, A) but never says what f computes or how simulation and verification feed into it §5.
• Lead-time guidance is extrapolated. The few-second lead time comes from takeover studies, not from studies of explanation delivery. The paper says time granularity of explanations is under-studied §5, so treat the conclusion as a hypothesis.
• Citation mismatch in tab. 2. The Teng et al. 2022 row cites the same key as Renz et al. (renz2022plant). The text cites Teng et al. under a different key (teng2022hierarchical). Check the bibliography entry before quoting it.
• Spelling variants in tables. Mankodiya is written as Mankondiya in tab. 3 and as Mankodiya in tab. 1 and the text, with the same citation key.
• Audience groups do not match. Section 3.3 defines four stakeholder groups (road users, AV developers, regulators and insurers, executives). The target-audience columns also use "Backup drivers", which appears only in tab. 5, and "Road users", which appears in tab. 1 and tab. 5.
• Figure numbers are from source order. Figures are numbered here in the order the LaTeX source introduces them. Check the PDF before citing a figure number. Figures 1, 4, 5, 6 and 8 credit third-party sources, and none is reproduced here.
• Version. The raw metadata records arXiv v5 (updated 2024-04-25). The card follows the LaTeX source that wikillm fetch retrieved, which is what the citations refer to. An earlier arXiv version may read differently.
• License. arXiv lists CC BY 4.0. Prose is written in our own words.

Lineage

• Builds on: no method. The paper compares itself with earlier surveys: Omeiza et al., a conceptual XAI framework for modular driving; Zablocki et al., an overview of end-to-end vision-based driving; and Kuznietsov et al., a systematic review that proposes the SafeX framework §4.1.
• Own earlier work named by the authors: Atakishiyev et al., on end-to-end learning for AVs §5; Atakishiyev et al. (2023), on VQA explanations for driving actions tab. 1 §4.3; and Atakishiyev et al. (2024), on the robustness of video QA models to adversarial questions tab. 6 §6.
• Compared against: not applicable, since the paper has no experiments.
• Superseded by: not known from this wiki.