Glass-Box Automated Driving: Insights and Future Trends

Mauro Bellone

1 a

, Raivo Sell

2 b

and Ralf-Martin Soe

1 c

FinEst Centre for Smart Cities, Tallinn University of Technology, Estonia

Department of Mechanical and Industrial Engineering, Tallinn University of Technology, Estonia

Keywords:

Glass-Box Models, Automated Driving, Performance vs Interpretability Trade-off.

Abstract:

Automated driving has advanced signiﬁcantly through the use of black-box AI models, particularly in percep-

tion tasks. However, as these models have grown, concerns over the loss of explainability and interpretability

have emerged, prompting a demand for creating ’glass-box’ models. Glass-box models in automated driving

aim to design AI systems that are transparent, interpretable, and explainable. While such models are essential

for understanding how machines operate, achieving perfect transparency in complex systems like autonomous

driving may not be entirely practicable nor feasible. This paper explores arguments on both sides, suggesting

a shift of the focus towards balancing interpretability and performance rather than considering them as con-

ﬂicting concepts.

1 INTRODUCTION

Fully automated driving on roads has been a long-

sought goal, with little signiﬁcant progress for many

years (Stanton and Young, 1998). The ﬁeld has only

recently advanced notably, driven by improvements in

hardware computational capabilities and data-driven

models with the promise of end-to-end automated

driving in the near future (Yurtsever et al., 2020).

However, this progress comes at the cost of losing a

clear connection to the fundamentals of process con-

trol (Omeiza et al., 2021). Automated driving on

roads exempliﬁes the control of a complex system

where data-driven models offer advantages, as creat-

ing detailed analytical models for every component is

nearly infeasible.

Classical optimal control theory aims to design

control processes that achieve optimal tracking of a

desired reference signal, offering elegant analytical

solutions to many problems. The theory begins with

linear, scalar systems and extends to multivariate,

non-linear processes, signiﬁcantly increasing compu-

tational complexity and the system’s level of abstrac-

tion, particularly due to the challenges of solving

higher-dimensional and non-linear models.

From this perspective, a model that analytically

captures reality at its fundamental level would be

the ideal foundation for building deterministic, error-

https://orcid.org/0000-0003-3692-0688

https://orcid.org/0000-0003-1409-0206

https://orcid.org/0000-0002-6782-1677

proof control systems. By deﬁnition, such a system

would qualify as a glass-box model, offering com-

plete interpretability and explainability. On the other

hand, many fundamental physical processes remain

poorly understood, and while the human instinct is to

seek clear explanations, technological advancement

often relies on approximate descriptions to manage

complex control systems in an uncertain real world.

A deeper understanding leads to improved function-

ality, highlighting the importance of interpretabil-

ity. While black-box models may enable control sys-

tems to function effectively, only full analytical inter-

pretability can unlock their true potential.

Following the example of automated driving, a

full analytical and computational description of dy-

namic behavior—such as aerodynamics, tire friction,

and engine response to driver commands—is com-

plex and not entirely practical yet achievable. Steer-

ing robots capable of driving vehicles in structured

environments, such as test tracks, have existed since

the 1980s (Weisser et al., 1999). The real challenge,

however, lies in real-world interactions, where model-

ing the unpredictable behavior of other road users be-

comes increasingly difﬁcult. In such scenarios, users

expect a robotic driver to act deterministically and

make safety-critical decisions within fractions of a

second.

Following this line of research, this work dis-

cusses the following research questions while focus-

ing on the problem of automated driving:

RQ1. What are the practical motivations for build-

ing glass-box analytical models?

880

Bellone, M., Sell, R. and Soe, R.-M.

Glass-box Automated Driving: Insights and Future Trends.

DOI: 10.5220/0013384300003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 880-885

ISBN: 978-989-758-737-5; ISSN: 2184-433X

RQ2. What are the challenges in achieving full ex-

plainability and interpretability in the ﬁeld of

automated driving?

RQ3. What viable strategies exist for achieving

transparency in automated driving systems?

This paper is structured as follows: Section 2 pro-

vides the motivation for applying glass-box models

and presents our perspective on RQ1. Section 3 ex-

plores the challenges of developing such models in

the ﬁeld of automated driving, addressing RQ2. Fi-

nally, our proposed strategy for RQ3, which focuses

on modularizing models and balancing interpretabil-

ity with performance, is discussed in Section 4.

2 MOTIVATION

Let’s assume we want to design a system capable of

taking the best action in every situation, following a

data-driven approach. For simplicity, we can envi-

sion it as a perfect driver: one that always selects the

optimal route from A to B while ensuring energy ef-

ﬁciency, minimal travel time, safety, and passenger

comfort.

The process of building such a system would in-

volve recording pairs of information points and opti-

mal actions. Each information point can be consid-

ered as an abstract representation of everything the

robotic driver perceives, including the internal vehi-

cle status and any external sources of information

(e.g., driving scenario, trafﬁc conditions, other road

users, etc.). For simplicity Fig. 1 provides an ab-

stract representation of these information points on an

action-information Cartesian map, where i represents

our information content and d represents our best

recorded decision in such a scenario. By consider-

ing all recorded information, represented as gray dots,

one can construct a function—using any analytical

ﬁtting method—thus obtaining an action-information

mapping function.

The goal of building such a function is to use the

action-information model as a decision-making tool

to estimate the best action as new information be-

comes available. If we assume that the blue line in

Fig. 1 represents our best model, when new informa-

tion i

new

arrives, the model’s guess should be no other

than

new

as the intersection point between the infor-

mation and the model. However, the optimal decision

in that situation may be something slightly different,

such as d

∗

new

, resulting in a gap between the model’s

guess and the true optimal decision.

Such an gap in the optimality might result from:

1. Measurement errors in both the information used

to build the model and the new data point.

2. Approximation errors due to complex interactions

between variables not fully captured by the model.

3. Truly unknown circumstances, highlighting a dif-

ference between the current information point and

previous knowledge of the system.

4. A change in the operational domain, where the

model was not trained or calibrated for new con-

ditions, leading to discrepancies in the decision-

making process.

Sticking to the example of automated driving, the

gap in the optimal decision can lead to a wide range

of consequences, from minor delays in travel time or

passenger discomfort to more severe outcomes like

crashes. How can one identify the cause of a miscal-

culation in the decision-making process if the system

is a fully enclosed black box?

From this perspective, all points mentioned ear-

lier—measurement errors, approximation errors, truly

unknown circumstances, and changes in the opera-

tional domain—are equally valid sources of error with

very little possible action to address and correct such

errors to ensure the system functions properly across

different scenarios. The current black-box approach

to solving this problem tends to add information to the

system indeﬁnitely, leading to memorization rather

than meaningful knowledge abstraction with minimal

validation and veriﬁcation possibilities (Pikner et al.,

2024).

Assuming that our action-information model is an

analytical multivariate function underpinning our pro-

cess, a gap in decision-making can be investigated,

studied, and potentially debugged if, and only if, the

system can be fully explained and interpreted.

Often, interpretability and explainability are seen

as a written form of text providing an explanatory

Figure 1: Action - information model depiction. Gray dots

constitute information points used to generate the ﬁtting

function in blue. The optimality gap arise from the dif-

ference between the optimal decision d

∗

new

and the model

guess

new

Glass-box Automated Driving: Insights and Future Trends

881

statement about driving decisions and speciﬁc actions

(Omeiza et al., 2022). However, this approach should

not be conﬂated with true interpretability or account-

ability. The generated text, often produced by black-

box LLMs (large language models), merely describes

the situation without offering useful debugging infor-

mation. As LLMs have shown, it is always possible

to generate ‘human-understandable’ explanations that

can be terribly misleading and fail to reﬂect the sys-

tem’s actual behavior at its fundamental level. Thus,

this type of verbal explanation often falls more into

the realm of psychological perspective rather than a

genuine interpretation of system behavior.

3 CHALLENGES

Without any doubt, full black-box models have

demonstrated compelling capabilities in controlling

complex systems such as autonomous vehicles (Chen

et al., 2024). This achievement cannot be neglected,

as it represents a signiﬁcant opportunity to address

long-term practical problems. However, glass-box

models offer the opportunity to debug and reﬁne these

systems by providing full control and transparency.

This allows for a deeper understanding and improve-

ment of each system’s behavior (Kuznietsov et al.,

2024).

The ﬁrst signiﬁcant challenge in building a glass-

box control system for autonomous driving lies in

managing the complexity of the environment. As pre-

viously mentioned, the challenge does not stem from

driving along a predeﬁned path at a precise velocity

but rather from interacting effectively with the sur-

rounding environment. Autonomous driving requires

processing vast amounts of real-time data from sen-

sors such as cameras, LiDAR, radar, and GPS. The

decision-making process entails intricate interactions

between perception, prediction, and control systems,

which are often modeled using complex deep learning

or neural networks—methods that are inherently dif-

ﬁcult to interpret. Figure 2 illustrates the control ﬂow

from low-level to high-level driving models. The con-

cept is that any state-of-the-art controller, from PID

(Emirler et al., 2014) to Lyapunov-based controllers

(Alcala et al., 2018) (Karafyllis et al., 2022), can

effectively and precisely manage vehicle speed and

steering angle, provided a simpliﬁed physical model

of the vehicle is known. The assumptions required

for these controllers to function are often unrealis-

tic and ineffective at predicting dynamic environmen-

tal changes. The higher the complexity of the ve-

hicle model, the better and more effective the con-

troller. This capability is sufﬁcient for driving in

structured environments. The system’s internal com-

ponents, highlighted with a dashed line in Fig. 2, can

be fully explainable and interpretable PIDs. Litera-

ture offers a robust foundation of analytical analyses

and formal solutions for the problem of following ref-

erence signals, enabling deterministic control in these

scenarios (Fleming and Rishel, 2012).

As the research community approached this prob-

lem, it quickly became evident that addressing it re-

quired multiple levels of abstraction. At the very low

level, classical automatic controllers, such as PID or

model predictive controllers, perform their tasks ef-

fectively. However, tackling vehicle perception and

routing with the same level of detail as low-level con-

trol, represented in the outer cycle of Fig. 2, is com-

putationally and practically infeasible.

The routing problem itself can be subdivided into

two distinct components: the generation of kinemat-

ically or dynamically feasible trajectories for short-

range vehicle control, and waypoint generation for

high-level vehicle routing. The former requires solv-

ing differential equations, which are computationally

demanding and often limited to localized regions.

The latter typically employs simpler, non-physically

compliant algorithms, such as Dijkstra’s algorithm

or rapidly-exploring random trees (RRT) (LaValle,

2006). Attempting to solve differential equations for

every point along a long path is not only computa-

tionally prohibitive but also ineffective, as it assumes

a static environment, thereby reducing the system’s

ability to adapt to dynamic changes. Conversely, ig-

noring differential equations altogether can result in

unfeasible paths that violate the vehicle’s physical

constraints.

The routing problem exempliﬁes the necessity

for different levels of abstraction to address com-

plex challenges in autonomous driving. Each level

retains its own interpretability: long-range planning

may sacriﬁce physical precision but provides compu-

tational efﬁciency, while short-range trajectory gen-

eration maintains a detailed physical interpretation.

This balance allows the system to adapt to a dynamic

environment while ensuring feasibility at a local level.

On a different level, the perception problem,

which includes tasks such as object detection and seg-

mentation, is predominantly addressed using black-

box models (Huang et al., 2022). On one hand, these

models excel in providing detection capabilities that

can even surpass human performance in certain sce-

narios, such as low-light conditions or when integrat-

ing data from multiple sensor sources. An illustrative

example is shown in the two images in Fig. 2 which

are extracted from the IseAuto dataset (Gu et al.,

2023) and include the addition of a black-box seg-

IAI 2025 - Special Session on Interpretable Artiﬁcial Intelligence Through Glass-Box Models

882

Figure 2: Levels of control abstractions featuring low-level

process control and high-level intelligent functions. Visual

images from IseAuto dataset (Gu et al., 2023).

mentation result (Gu et al., 2024).

On the other hand, they lack an understanding of

the semantic meaning of objects within a scene, treat-

ing a tree and a person merely as objects without rec-

ognizing their distinct roles or contextual importance.

Contextual importance is inherently challenging to

capture, as even human observers may assign vary-

ing interpretations and signiﬁcance to similar objects

in a scene. An interpretable and explainable model

should not aim to produce verbose descriptions of a

scene but rather to classify objects accurately while

accounting for their high-level roles and importance.

For instance, in such a framework, errors or ﬂaws in

detecting vehicles or pedestrians, compared to non-

critical objects like rocks or foliage, could be identi-

ﬁed, debugged, and addressed in a targeted manner.

3.1 Safety-Critical Nature

While it may seem that autonomous vehicles must pri-

oritize split-second decision-making over immediate

explanations, it is precisely in these high-stakes sce-

narios that the effort to build simple, classical control,

and rule-based systems becomes essential. Trans-

parency is critical for diagnostics and regulatory com-

pliance, and it is often mischaracterized as compro-

mising the vehicle’s ability to respond effectively in

emergencies. On the contrary, such transparency en-

sures that misunderstandings or misinterpretations are

minimized, particularly in situations where they can-

not be tolerated. In (Abrecht et al., 2024), the au-

thors clearly emphasize the safety concerns that deep

learning poses to automated driving, covering aspects

such as operational domain deﬁnition and limitations,

as well as the methods used for data preparation and

algorithm development. Moreover, regulations are

placing increasing emphasis on the importance of AI

explainability in safety-critical industries like trans-

portation. Glass-box models are required to comply

with stringent industry standards, such as ISO 26262

(ISO, 2011) (functional safety) and ISO/PAS 21448

(Safety of the Intended Functionality - SOTIF), to en-

sure reliability, accountability, and safety (Kirovskii

and Gorelov, 2019). Additionally, frameworks such

as the European Union’s Ethics Guidelines for Trust-

worthy AI and the U.S. NIST’s initiatives on AI ex-

plainability advocate for more transparent and ac-

countable AI systems. The recently adopted AI Act

(Regulation (EU) 2024/1689 laying down harmonized

rules on artiﬁcial intelligence) sets clear requirements

and obligations for AI developers, emphasizing ex-

plainability and accountability in AI-based systems,

including automated driving functionalities.

4 STRATEGIES AND FUTURE

TREND

Several strategies exist to achieve transparency, ac-

countability, and explainability in automated driving,

many of which can be applied to other complex sys-

tems that beneﬁt from black-box models. While a

fully glass-box model for automated driving may be

impractical in its purest form—particularly in com-

plex real-world scenarios—it remains an ambitious

long-term goal. In the interim, hybrid models rep-

resent a practical solution, leveraging the strengths

of both interpretable and black-box approaches. A

promising strategy is to combine interpretable models

for high-level decision-making with black-box mod-

els for perception tasks, striking a balance between

interpretability and efﬁcacy with the goal of reducing

black-box models to the miminum level. For exam-

ple, rule-based logic can be applied to lane-change

policies (Malayjerdi et al., 2022), providing clear

and explainable decision-making, while deep learn-

ing models handle object detection, which often bene-

ﬁts from the data-driven adaptability of black-box ap-

proaches

The application of simpler, domain-speciﬁc mod-

els also offers notable strengths in achieving trans-

parency and reliability. In constrained environments,

such as autonomous shuttles or last-mile delivery ve-

hicles, rule-based systems or simpler machine learn-

ing models can effectively balance explainability with

functionality. This approach aligns with the concept

of maintaining a human-in-the-loop framework, al-

lowing human operators to oversee decisions and in-

tervene when necessary. This ensures that actions

taken by the system adhere to ethical and safety con-

siderations. Moreover, adopting a modular system

design—breaking the driving stack into smaller, in-

terpretable modules—further aids in achieving trans-

parency. For example:

Glass-box Automated Driving: Insights and Future Trends

883

• Perception: Explains how objects are detected and

classiﬁed.

• Planning: Provides insights into why a particular

trajectory or decision was selected.

• Control: Demonstrates how the vehicle executes

the planned commands.

These modular explanations help ensure that the

system remains interpretable and debuggable, while

still beneﬁting from advancements in AI and automa-

tion. Even in such scenarios, designers must rely on

extensive simulation environments and formal veriﬁ-

cation methods to understand and validate system be-

haviors under diverse conditions.

Autonomous systems operate in the physical

world, a domain governed by the principles of physics

(e.g., Maxwell’s equations, Newton’s laws). These

principles provide a robust foundation for deﬁning

a governing framework that can enhance both ex-

plainability and interpretability. A critical question

emerges:

How might one leverage the fundamental prop-

erties of physics to build a validation governor

around AI-based autonomy systems?

This question remains an open avenue for re-

search, challenging the community to explore inno-

vative ways to integrate physical laws into the vali-

dation and explainability frameworks of autonomous

systems (Pikner et al., 2024).

The application of Post-Hoc Explainability such

as saliency maps, SHAP (Shapley Additive Explana-

tions) (Lundberg and Lee, 2017), (Lundberg et al.,

2020), or LIME (Local Interpretable Model-Agnostic

Explanations) (Ribeiro et al., 2016) to analyze and ex-

plain decisions made by black-box models constitute

also emergent solutions. One might note that Post-

hoc methods do not require the model to be inher-

ently interpretable but rather attempt to interpret and

explain the outputs of a AI model after it has been

trained, which makes them practical for use with com-

plex models that prioritize performance but sacriﬁce

interpretability.

5 CONCLUSION

A fully glass-box model for automated driving is

likely impractical in its purest form, particularly in

complex real-world scenarios as highly interpretable

models (e.g., decision trees or rule-based systems)

may struggle to capture the nuanced decision-making

required in dynamic driving environments. How-

ever, hybrid approaches that blend interpretability

with black-box models offer a practical way forward

while keeping the aspiration for full glass-box designs

alive. As is often the case in engineering, the optimal

solution lies in the middle ground—making the crit-

ical components of the system transparent enough to

ensure safety, reliability, and accountability, without

compromising performance.

The drive to pursue glass-box models is deeply

rooted in the fundamental curiosity of engineers. The

question, ’Why does a system work the way it does?’

forms the very basis of scientiﬁc exploration. With-

out such curiosity, humanity might still accept a ﬂat

Earth as truth, never challenging experiences that

seem counterintuitive or seeking alternative perspec-

tives. The essence of progress lies in questioning and

reframing our understanding of the world and its com-

plexities.

ACKNOWLEDGMENT

Part of this research has received funding from the Eu-

ropean Union’s Horizon 2020 Research and Innova-

tion Programme under grant agreement No. 856602

(Finest Twins) and from the European Union’s Hori-

zon Europe Research and Innovation Programme un-

der grant agreement No. 101135988 (PLIADES:

AI-Enabled Data Lifecycles Optimization and Data

Spaces Integration for Increased Efﬁciency and Inter-

operability)

REFERENCES

Abrecht, S., Hirsch, A., Raafatnia, S., and Woehrle, M.

(2024). Deep learning safety concerns in automated

driving perception. IEEE Transactions on Intelligent

Vehicles, pages 1–12.

Alcala, E., Puig, V., Quevedo, J., Escobet, T., and Comaso-

livas, R. (2018). Autonomous vehicle control using a

kinematic lyapunov-based technique with lqr-lmi tun-

ing. Control engineering practice, 73:1–12.

Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., and

Li, H. (2024). End-to-end autonomous driving: Chal-

lenges and frontiers. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 46(12):10164–

10183.

Emirler, M. T., Uygan,

I. M. C., Aksun G

uvenc¸, B.,

and G

uvenc¸, L. (2014). Robust pid steering con-

trol in parameter space for highly automated driv-

ing. International Journal of Vehicular Technology,

2014(1):259465.

Fleming, W. H. and Rishel, R. W. (2012). Deterministic

IAI 2025 - Special Session on Interpretable Artiﬁcial Intelligence Through Glass-Box Models

884

and stochastic optimal control, volume 1. Springer

Science & Business Media.

Gu, J., Bellone, M., Pivo

nka, T., and Sell, R. (2024).

Clft: Camera-lidar fusion transformer for semantic

segmentation in autonomous driving. arXiv preprint

arXiv:2404.17793.

Gu, J., Lind, A., Chhetri, T. R., Bellone, M., and Sell, R.

(2023). End-to-end multimodal sensor dataset col-

lection framework for autonomous vehicles. Sensors,

23(15):6783.

Huang, K., Shi, B., Li, X., Li, X., Huang, S., and

Li, Y. (2022). Multi-modal sensor fusion for

auto driving perception: A survey. arXiv preprint

arXiv:2202.02703.

ISO (2011). Iso 26262 road vehicles– functional safety. ISO

Standard (2011).

Karafyllis, I., Theodosis, D., and Papageorgiou, M. (2022).

Lyapunov-based two-dimensional cruise control of

autonomous vehicles on lane-free roads. Automatica,

145:110517.

Kirovskii, O. and Gorelov, V. (2019). Driver assistance sys-

tems: analysis, tests and the safety case. iso 26262 and

iso pas 21448. In IOP Conference Series: Materials

Science and Engineering, volume 534, page 012019.

IOP Publishing.

Kuznietsov, A., Gyevnar, B., Wang, C., Peters, S., and

Albrecht, S. V. (2024). Explainable ai for safe and

trustworthy autonomous driving: A systematic review.

arXiv preprint arXiv:2402.10086.

LaValle, S. M. (2006). Planning algorithms. Cambridge

university press.

Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin,

J. M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N.,

and Lee, S.-I. (2020). From local explanations to

global understanding with explainable ai for trees. Na-

ture machine intelligence, 2(1):56–67.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed ap-

proach to interpreting model predictions. In Guyon, I.,

Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R.,

Vishwanathan, S., and Garnett, R., editors, Advances

in Neural Information Processing Systems 30, pages

4765–4774. Curran Associates, Inc.

Malayjerdi, E., Sell, R., Malayjerdi, M., Udal, A., and Bel-

lone, M. (2022). Practical path planning techniques in

overtaking for autonomous shuttles. Journal of Field

Robotics, 39(4):410–425.

Omeiza, D., Web, H., Jirotka, M., and Kunze, L. (2021).

Towards accountability: Providing intelligible expla-

nations in autonomous driving. In 2021 IEEE Intelli-

gent Vehicles Symposium (IV), pages 231–237. IEEE.

Omeiza, D., Webb, H., Jirotka, M., and Kunze, L. (2022).

Explanations in autonomous driving: A survey. IEEE

Transactions on Intelligent Transportation Systems,

23(8):10142–10162.

Pikner, H., Malayjerdi, M., Bellone, M., Baykara, B. C.,

and Sell, R. (2024). Autonomous driving validation

and veriﬁcation using digital twins. VEHITS, pages

204–211.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ”why

should I trust you?”: Explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

International Conference on Knowledge Discovery

and Data Mining, San Francisco, CA, USA, August

13-17, 2016, pages 1135–1144.

Stanton, N. A. and Young, M. S. (1998). Vehicle automation

and driving performance. Ergonomics, 41(7):1014–

1028.

Weisser, H., Schulenberg, P., Gollinger, H., and Michler, T.

(1999). Autonomous driving on vehicle test tracks:

overview, implementation and vehicle diagnosis. In

Proceedings 199 IEEE/IEEJ/JSAI International Con-

ference on Intelligent Transportation Systems (Cat.

No. 99TH8383), pages 62–67. IEEE.

Yurtsever, E., Lambert, J., Carballo, A., and Takeda, K.

(2020). A survey of autonomous driving: Common

practices and emerging technologies. IEEE access,

8:58443–58469.

Glass-box Automated Driving: Insights and Future Trends

885