Decision-Making Techniques for Autonomous Vehicles

Decision-Making Techniques for Autonomous Vehicles provides a general overview of control and decision-making tools that could be used in autonomous vehicles. Motion prediction and planning tools are presented, along with the use of machine learning and adaptability to improve performance of algorithms in real scenarios. The book then examines how driver monitoring and behavior analysis are used produce comprehensive and predictable reactions in automated vehicles. The book ultimately covers regulatory and ethical issues to consider for implementing correct and robust decision-making. This book is for researchers as well as Masters and PhD students working with autonomous vehicles and decision algorithms.

• Provides a complete overview of decision-making and control techniques for autonomous vehicles
• Includes technical, physical, and mathematical explanations to provide knowledge for implementation of tools
• Features machine learning to improve performance of decision-making algorithms
• Shows how regulations and ethics influence the development and implementation of these algorithms in real scenarios

• Language: English
• Publisher: Elsevier Science
• Release date: Mar 3, 2023
• ISBN: 9780323985499

Book preview

Chapter 1: Overview

Jorge Villagraa; Felipe Jiménezb    a CSIC, Madrid, Spain

b Universidad Politécnica de Madrid, Madrid, Spain

Abstract

Although there has been an impressive progress of artificial decision-making systems, there are still major challenges ahead to produce safe and efficient outputs in more and more driving scenarios. This book contributes to understand the state of the technology in AVs and its interconnections with perception and control systems. It presents a comprehensive review of principles and methods covering both classical and modern techniques, including not only the possibilities offered by the on-board systems but also from the infrastructure. These technical aspects are moreover complemented with some deployment issues and insights into the user influence. This introductory chapter describes the scope and structure of the book, where the problems tackled by every technology are formalized and the existing techniques are classified and briefly described.

Keywords

Decision-making; Autonomous vehicles

1.1: Introduction

According to SAE (2021), Automated Vehicles (AVs) are equipped with any driving automation system capable of performing dynamic driving tasks on a sustained basis. From the user perspective, the ultimate goal of these vehicles would be to provide the ability to the passengers or goods to move from a starting to a finishing point in a safe and optimal manner, without human intervention. To make this dream come true, the traditional automotive value chain (OEM, Tier 1, Tier 2), research centers, and new technology players are investing heavily in several autonomy-related research fields. The interest behind such a move is related to the great transformation potential of the technology, especially in urban environments, and the need of solving the problems that have caused the exponential growth of road mobility.

Indeed, cities and metropolitan areas are the nerve centers of global growth and will reach 60% of the world's population by 2030 (UN, 2022) and more than 80% in the EU by 2050 (EC, 2019a). This increasing urbanization is generating disorderly growth, with inadequate and overburdened infrastructure and services, with freight and passenger transport being one of the clearest examples of inefficiency and negative impact on air pollution. Goal 11 of the UN SDGs (UN, 2022) has among its targets to provide access to safe, affordable, accessible, accessible, and sustainable transport systems for all and improve road safety, paying special attention to the needs of persons in vulnerable situations, such as women, children, persons with disabilities, and older persons.

In line with this objective, the European Commission recently signed the European Green Deal (EC, 2019a)—which highlights that road transport accounts for 18% of greenhouse gas emissions and 30% of particulate matter emissions in the EU—setting the achievement of sustainable transport as a priority. More specifically, it refers to the need to focus on users and provide them with more affordable, accessible, and healthier alternatives to their current mobility habits.

Avoiding the most negative effects of climate change means drastically reversing the current trend in the number of trips in cities—between 2010 and 2016 congestion increased in London by 14%, in Los Angeles by 36%, and in Paris by 9%. To achieve this, it is essential to rethink our major urban areas and orient them toward a zero-carbon future. Innovation in urban mobility has great potential to foster this imminent and necessary transformation.

Autonomous mobility can be an effective tool to pursue these ambitious environmental goals while increasing inclusiveness and equity in access. Indeed, some user segments do not have competitive alternative mobility options adapted to their needs—for example, people over 60 will make up a third of the population by 2050, so a flexible and efficient approach is needed to provide users with a real alternative to driving. If intelligently integrated with different forms of public transport, (shared) AVs can make a decisive contribution to improving the current negative aspects of urban mobility (congestion and pollution) by making it more affordable, efficient, user-friendly, and available to all.

According to traffic accident statistics, between 90% and 98% of traffic accidents are at least partly caused by humans, and of these, about 40% occur in urban environments (EC, 2019b). The introduction of autonomous driving, therefore, has great potential to reduce accident rates and should improve safety in most situations; however, in urban environments, where there is greater complexity (number of actors involved) and unpredictability (very different behaviors of pedestrians, micromobility users, all types of vehicles, etc.), the technology is not still mature enough to significantly reduce this risk of collision. Indeed, the roadmaps drawn up by the ERTRAC (Gräter et al., 2021) and the NHTSA (US NSTC, 2020) predict that we will see a gradual appearance of systems with increasing assistance capabilities, but none of them dares to estimate a precise date for the appearance of a system with full autonomy (in any context, including the most complex, urban). The reasons for this caution lie, on the one hand, in the important socio-economic aspects still to be resolved, but above all in some significant technological barriers.

Indeed, there is a major challenge not only to get as close as possible to the goal of zero accident victims, but also to be able to know in advance precisely the driving situations in which the risk of a collision increases considerably. And this predictability is a key factor to enhance users trust, probably the most relevant obstacle for the massive deployment of Avs. To become aware of the scope of that challenge, it is necessary to understand the essential operating mechanism of a vehicle that aspires to be autonomous. The artificial system must be continuously aware of its environment, first perceiving (processing all the data captured by sensors), then correctly structuring and prioritizing all the information obtained (world modeling or interpretation), and finally making decisions based on the interpretation of the scene, which have to be dependent on the driving context. Furthermore, the actions derived from these decisions have to be taken within fractions of a safe fraction of a second and ensure a safe behavior in the vehicle under all circumstances.

Although there has been an impressive progress in developing artificial decision-making systems operational in this context, producing safe and efficient outputs in more and more driving scenarios, there are still major challenges ahead. Indeed, the current state of the technology could make satisfactory decisions with different types of Advanced Driver Assistance Systems (ADAS), functional in quite restrictive domains, but is far from achieving full autonomy in any scenario. To have a better understanding of where and how these systems work today and what it is still required, the next section introduces the different automation levels, the notion of Operational Design Domain (ODD), and their influence in decision-making techniques.

1.2: Decision-making, automation levels, and operational design domains

The Society of Automotive Engineers (SAE) proposed some years ago a taxonomy of automated driving systems (ADS) ranging from level 0, with no driving automation, to level 5, full driving automation (SAE, 2021). To understand the main differences among them, the following four key parameters were specified:

•Execution of steering, acceleration, and deceleration

•Monitoring of the environment

•Fallback performance in any dynamic task

•System capability (none, some or all modes of automated driving)

Up to level 2, it is always the driver who is responsible for monitoring the environment, the main difference between levels 1 and 2 being the system capability (only one mode versus all modes). At level 3 and above, the ADS is responsible for monitoring the environment, but the driver needs to be capable of resuming control after a certain amount of time (still to be precisely characterized) in case an alert arises and a handover is requested. At level 4, requests to intervene are still possible, but no longer mandatory, meaning that the fallback performance does not include the human driver as a necessary component anymore. The main difference of level 4 with level 5 is that in the former, the ADS operates only in a limited area of operation (ODD), whereas in the latter, this area of operation is unrestricted.

According to SAE (2021), the definition of the ODD is as follows: operating conditions under which a given driving automation system or feature thereof is specifically designed to function, including, but not limited to, environmental, geographical, and time-of-day restrictions, and/or the requisite presence or absence of certain traffic or roadway characteristics.

Any ADS is composed of multiple functions or features, whose performance fulfills the design requirements for a given automation level and ODD. For instance, the Traffic Jam Assist is a system with an SAE Level 3 that can be used in situations with dense traffic and under reasonable weather and visibility conditions. However, the same vehicle could include Level 2 ADAS in wider ODDs (e.g., inter-state roads or even urban streets) or Level 4 in specific structured environments (e.g., valet chauffeur in parking lots). As a result, and until full autonomy is not reached, the capabilities in terms of decision-making can be significantly different depending on the autonomy level under consideration and the specific ODD where it is deployed. The next section is devoted to define the scope of the decision-making techniques and tools discussed in this book.

1.3: Scope of the book

There are a number of interesting books in the literature dealing with the main technologies used in intelligent vehicles (e.g., Jimenez, 2017; Watzenig and Horn, 2016). Some of them provide quite high-level descriptions of architectures and components, and some others focus on a more specific technology brick (e.g., perception), and when it comes to decision-making, mostly planning (LaValle, 2006) and control (Kiencke and Nielsen, 2000) techniques are at the heart of the existing monographies. However, works covering the whole processing pipeline between the world modeling and low-level control with a broad look are lacking. Besides, the sense-think-act robotics paradigm, an undeniable source of inspiration for the automated driving community, has been lately stirred by the growing interest of AI-enabled approaches. A comprehensive review of principles and methods covering both classical and modern techniques, including not only the possibilities offered by the on-board sensors but also from the infrastructure, was needed and is the motivation for this book.

As it is not possible to describe in depth all the existing knowledge body (several hundreds of works are published every year), we aimed to focus on formalizing the problems tackled by every technology, providing thereafter taxonomies and brief descriptions of the existing techniques and including as many references as possible for the reader to deepen in the subjects of his/her interest. In other words, we aspire at positioning as a reference work with which engineers and researchers in their early years can identify the existing decision-making approaches and interactions among components.

For clarity, Fig. 1.1 shows a generic architecture of an AV. Five main elements can be appreciated: localization, perception, decision-making, interaction, and control. The main question each one of these technology bricks deals with is as follows: Where am I? Where can I move? How should I move? How do I interact? and How do I act? We assume that some answers exist to the first two questions (beyond the scope of the book), provide solutions to the third and fourth problems, and describe the main elements to be considered in the final processing stage, the low-level control, also out of the scope of the book.

Fig. 1.1

Fig. 1.1 General architecture of an autonomous vehicle.

It should be noted that the decision is not limited to arbitrate among a discrete number of possibilities following a number of criteria (safety, efficiency, ethics, regulations, comfort, consumption, driver preferences, etc.). It goes beyond some sort of multicriterion optimization problem, as it does not only select the best tactics on a given driving scenario, but it also thereafter produces a motion plan considering the most likely evolution of the surrounding road agents.

Before elaborating on the specific hypotheses considered in the book, a proper definition of the decision is provided, as the term can be understood from different viewpoints. Michon (1985) proposed to structure any robotics decision architecture around three main abstraction layers: strategic (also known as routing), tactical (or maneuvering), and operational (named also stabilization or control). Although there will be some mentions to route planning and to the interplay with low-level control components, the focus of the book will undeniably be in tactics: maneuvering (behavior) and motion planning, considering any technical or user-related aspect that may affect or replace this system.

The following list of features helps define the baseline from which the writing processed has been structured:

•It covers decision-making technologies for Connected and Automated Vehicles (CAVs) for on-road navigation. It does not tackle, therefore, the specificities of any other type of autonomous transport system (aerial, marine, underwater or even off-road ground).

•The technology described in the book is oriented toward full automation; however, applications targeting SAE Levels 3 and 4 are also considered in some chapters.

•As already mentioned, AVs need perception and localization systems to make the most adapted decision in every context. As a result, it is assumed in the sequel that these systems exist and can inform about the pose of the vehicle in the world and the drivable space around it so that a preliminary situational awareness is provided to the system. In other words, a world model has to be provided to the decision-making system. This environment representation can be fed by different types of sensors: exteroceptive (LiDAR, radar, cameras, and ultrasounds) and proprioceptive (GNSS, IMU, and encoders). In addition to that, it is assumed that some sort of digital map exists so that route planning and scene interpretation can be properly handled.

•Uncertainty management is one of the greatest challenges of decision-making systems. It appears in the shape of sensor's noise or inaccuracies, in perception occlusions, or misclassifications, leading to wrong road agent's pose estimation or motion prediction, resulting, in turn, to an inappropriate vehicle's surrounding understanding. The influence of this uncertainty and the techniques that explicitly take it into consideration are described in the book.

•Many existing works in behavior planning, motion prediction, motion planning, and even control are ODD-oriented. In the description of each method or technique, references to applications for specific layouts (e.g., intersections and highways) or scene configurations (presence of multiple road agents, different vehicle behaviors, and occlusions) are given.

•The most complex driving scenarios involve many interacting road users, often occluded behind them, that becomes extremely challenging for the AV to produce efficient and even safe decisions. As a result, the role of the infrastructure in embedded decision-making can be critical. The book does not cover decision algorithms for traffic management systems, but it provides interesting insights into (i) the influence of traffic-dependent routing on embodied tactical decisions; (ii) the cooperative possibilities of V2X communication technologies to better understand the environment dynamics; (iii) the way road infrastructure should be re-engineered to contribute to more predictable and safer actions.

•The liability and ethical implications of decisions are key in many areas of our life in society. The delegation of these responsibilities on AI-based machines requires to setup a clear safety framework toward high levels of user trust. To that end, rules and standards are under construction covering three major pillars: legal regulation, ethical considerations, and procedure for system validation and certification. The book also provides hints on the latest advances in these directions.

•Although the academic research has settled the theoretical bases of most of the technology bricks involved in decision-making in the last two decades, there has been a decisive and increasing involvement of the industry to transform these methods into viable products. The content of this book is the result of the collaboration of experts from different fields, fed in turn by public research literature. It should be noted that many of the presented strategies are today part of the most advanced prototypes of companies developing pilot services in restricted urban areas; however, implementation details of these latest automated driving functions are either confidential or patented, and therefore only vaguely covered in the book.

1.4: Book structure overview

As can be observed in Fig. 1.2, the book is structured around four main blocks of chapters with different inter-connections:

•Embedded decision components: this block includes the techniques and tools required to produce intelligent, predictable, safe, efficient, and comfortable decision and control actions using perception and localization data. This is the more substantive section of the book and details all the on-board components to be integrated within an automated vehicle. It should be noted that four chapters are connected in a dataflow that starts in the route planning system and ends with the interplay between decision and control. They follow the actual processing pipeline on-board the AVs and are encircled in a dotted box, together with an overarching chapter describing the main families of embodied decision architectures. As alternative AI-powered, end-to-end architectures are gaining interest, a dedicated chapter is included in this block, somehow keeping apart from the more traditional decision paradigms.

•Infrastructure-oriented decision-making: this block includes elements that can influence the on-board decision making from the infrastructure, either with traffic-dependent routing systems, with cooperative functions for a safer driving or providing guidelines for a more adapted infrastructure to the existing and oncoming AV's decision algorithms.

•User influence: this block is related to the interaction between the automated system and the driver/passenger and naturally connects with the vehicle control at the operational and tactical levels. It deals, on the one hand, with algorithmic strategies for shared control and, on the other hand, with psychology-based behavior models of the driver, useful to properly understand the handover possibilities of humans in every driving context.

•Deployment issues: this block includes transversal aspects that do not influence individually any specific building block, but rather the whole decision architecture. On one hand, they cover decision-related legal, ethical, and acceptance issues and, on the other hand, some indications on the ongoing initiatives to systematize the Automated Driving Functions (ADF) validation process, aiming at generating safety-oriented standards that will constitute a lever for trust.

Fig. 1.2

Fig. 1.2 Scheme depicting the book structure and the interconnections among chapters.

It should be noted that the first two blocks include fully technology-related chapters, whereas the two others, although including in some cases specific algorithms and tools, have clear connections either with the human side (acceptance, interactions) or with the applicability of the technologies in the real world (legal issues and validation aspects). A brief overview of the content included in each chapter is given below:

Chapter 2: Embodied decision architectures: This chapter introduces the specificities of the decision process in an embodied application (such as AVs). The way the different components are orchestrated has received an important inspiration from the robotics community, lately complemented with cognition-inspired approaches. In parallel, safety is a fundamental consideration in any transport-oriented AI-enabled system, where human-in-the-loop mechanisms are expected to operate in a long transition period. As a result, different approaches have emerged with the ambition to put together these different initiatives. This chapter aspires to briefly introduce them, highlighting their specificities, advantages, and drawbacks.

Chapter 3: Behavior planning: Besides routing, the higher-level component in a hierarchical decision architecture is behavior planning, responsible of tactical moves, often related to the identification of the most appropriate maneuver in a given driving situation. This chapter introduces the basics of decision theory under uncertainty and in the presence of multiple agents. Once the theoretical foundations are settled, different families of techniques are presented. These approaches focus on one or several of the challenges related to behavior/planning separation: maneuver feasibility conflicts, environment topology handling, the generation of efficient and safe human-like maneuvers, and scalability of scenario-specific solutions.

Chapter 4: Motion prediction and risk assessment: The short-term identified behavior is extremely dependent on the situation understanding, which in turn is highly influenced by the motion prediction of the surrounding road agents (vehicles, pedestrians, bicyclists, etc.). This chapter introduces the different existing techniques in this connection, differentiating the problems of driver trait estimation, intention estimation, motion prediction, and risk assessment. They all try to provide useful information dealing with three main challenges: (i) the unobservability and variability of cognitive processes that govern human decision-making are inherent; (ii) the existence of complex interactions among drivers or other road users; (iii) the variety and complexity of road scenes. It should be noted that although the evolution of vulnerable road users is probably the hardest to precisely predict, the focus of this chapter will be in vehicles, as the literature in pedestrians forecast is closer to perception techniques.

Chapter 5: Motion search space: Once the temporal dimension is included in the work modeling, and in parallel to behavior planning, the estimation of the available motion search space is key for most of motion planning techniques. In this chapter, these algorithms are categorized and examined under different perspectives.

Chapter 6: Motion planning: The final stage to generate a precise plan for vehicle controllers is motion planning, often decomposed into path and speed planning. It is fed by the outputs generated at the subsystems described in the previous three chapters (behavior planning, motion prediction, and motion search space). This chapter formulates the problem and then provides a complete review of the most widely used approaches. First, path planning techniques are introduced, inheriting many robotics contributions along the years and including the algorithms that exploit the highly structured nature of road networks. They all seek (quasi-)complete strategies that look for the best trade-off between safety, comfort, and utility. Then, methods to generate appropriate speed profiles form arbitrary paths are presented. Finally, some approaches aiming at obtaining a joint path/speed optimization are reviewed.

Chapter 7: End-to-end architectures: The classic modular pipeline is widely adopted in the industry due to its great interpretability and stability, but alternative end-to-end paradigms are gaining interest due to their simplicity supported and the rise of deep learning. Although there these methods still lack fail to generalize to unseen environments, a dedicated chapter has been included to describe its bases, main approaches, and potentialities. Thus, imitation learning, reinforcement learning, and transfer learning are described in this context, where the sensor data are used in a single step to produce control actions.

Chapter 8: Interplay between decision and control: Although the scope of the book does not include control techniques and methods to guide a vehicle along a path and speed reference, this chapter evokes several relevant inter-connections between stabilization techniques and decision-making strategies. Some of the most relevant research questions analyzed in the chapter are (i) if planning and control should be separated or seamlessly integrated; (ii) the available control/decision inter-connecting architectures if all the actuators can be taken into consideration as a whole; (iii) the best stabilization paradigm (if it exists); (iv) the relation between control specifications and vehicle behavior constraints or ethical considerations.

Chapter 9: Traffic data planning and route planning: This chapter provides an overview on two inter-related fields with potential to influence on-board decision-making of automated vehicles: route planning and traffic prediction. The most relevant routing problems are introduced, followed by some key algorithms providing interesting results. The influence of traffic prediction in these methods is thereafter emphasized, and some prospective guidelines are finally provided toward the deployment of these techniques in the context of CAV.

Chapter 10: Cooperative driving: The limitations of on-board sensor data in complex driving scenarios can lead to wrong or even unsafe decisions. The complexity behind some highly uncertain and dynamic driving scenarios (e.g., intersections, merging, and roundabouts) suggests full automation may have usage limits in the near future. Cooperative services relying on Vehicle-to-everything (V2X) communications can contribute to have a completer and more reliable situation understanding. This chapter summarizes the state of this technology, reviewing the existing and forthcoming standards (from Day-1 to Day-3+) and their interaction with decision-making (a bit at the infrastructure level, but overall at the vehicle level).

Chapter 11: Infrastructure impact: This chapter analyzes a somehow neglected aspect by CAV technology developers. The current road network has been designed to be used by human-driven vehicles. Although CAVs are being designed to mimic humans with a safety layer, some specific issues derived from the sensors’ locations, significantly different than human eyes, may lead to undesired behaviors in disengagements. In this context, the study analyzes the role and impact of the infrastructure from several perspectives. After this analysis, different Smart Road Levels (SRLs) are proposed to characterize the suitability of the road stretch under consideration to the safe deployment of AVs. These levels, bounded by a minimal Risk Condition, will help not only Road Administrations and Operators but also automated vehicles so that the on-board decision-making can safely accommodate the specificities of the road under consideration.

Chapter 12: Driver behavior: As mentioned in the previous chapter, SAE Level 3 vehicles may still require disengagements (requests from the machine to the human to take over control). This chapter reviews current human-centered models to properly support the authority transitions in these vehicles with intermediate levels of automation. Indeed, human-automation requires cognitive-emotional models to predict and understand the human behavior in handovers. The response and acceptance of human passengers to this technology and the related vehicle design is also analyzed to ensure maximum comfort and usability.

Chapter 13: Human-machine interaction: As a complement to the previous chapter, this other one introduces the algorithmic strategies to safely articulate shared and traded control, using among other inputs, the psychology-based driver models described above. Shared control emphasizes the real-time cooperation at the control level between driver and automation, with a dynamic allocation of control authority. Traded control looks for a dynamic shift of the human role between the driver and passenger, with a variable level of automation according to the complexity of the driving scenario. This chapter provides and introduces both strategies, including recent developments in terms of frameworks and algorithms.

Chapter 14: Algorithms validation: As CAVs have to remain operational in an extremely wide range of operation, the required open road testing to reach at least critical crash rates is clearly prohibitive. This chapter presents the recent initiatives to cope with this reality, namely, (i) validation methodologies involving not only brute force testing on open roads or closed course testing sites but also different types of simulation-based platforms and combined real-virtual evaluation mechanisms; and (ii) safety frameworks, relying on novel standards that take into consideration both the scenario complexity and the adaptive behavior of the novel and upcoming AI-enabled decision functions.

Chapter 15: Legal and social aspects: Relevant technical challenges are ahead of us before decision-making systems can be massively deployed in any ODD; however, social and legal considerations are at least equally critical for a successful adoption of the technology. In this chapter, the liability issues related to artificial decision-making are described and compared among the most relevant European countries. In addition to that, as ethics is an inescapable consideration in any human-in-the-loop decision-making process, the most relevant ethics dilemmas are recalled, and some approaches to face them are provided. Finally, the key trigger for AVs to be on the roads is user acceptance. The different factors affecting the willingness to experience and use this technology are described also in this chapter.

References

EC. The European Green Deal. Brussels: European Commission, COM; 2019.640 (12/2019).

EC. Annual Accident Report 2018. European Road Safety Observatory, European Commission; 2019.

Gräter A., Steiger E., Harrer M., Rosenquist M. Connected, Cooperative and Automated Driving—Update of ERTRAC Roadmap. ERTRAC Working Group, Brussels, Belgium, Tech. Rep; 2021.9.

Jimenez, 2017 Jimenez F., ed. Intelligent Vehicles: Enabling Technologies and Future Developments. Butterworth-Heinemann; 2017.

Kiencke U., Nielsen L. Automotive Control Systems: For Engine, Driveline, and Vehicle. Berlin, Heidelberg: Springer; 2000.

LaValle S.M. Planning Algorithms. Cambridge University Press; 2006.

Michon J.A. A critical view of driver behavior models: What do we know, what should we do?. In: Human Behavior and Traffic Safety. Springer; 1985:485–524.

SAE. Taxonomy and definitions for terms related to on-road motor vehicle automated driving systems. SAE Stand. J. 2021;3016:1–16.

UN. https://www.un.org/sustainabledevelopment/. 2022.

US NSTC. Ensuring American Leadership in Automated Vehicle Technologies: Automated Vehicles 4.0. Las Vegas: Recuperado el; . 2020;vol. 25 (2020-02).

Watzenig and Horn, 2016 Watzenig D., Horn M., eds. Automated Driving: Safer and more Efficient Future Driving. Springer; 2016.

Part I

Embedded decision components

Chapter 2: Embodied decision architectures

Jorge Villagra; Antonio Artuñedo    CSIC, Madrid, Spain

Abstract

The decision system architecture of an automated vehicle plays a crucial role in its performance, as has been evidenced in the existing prototypes deployed in real-life driving scenarios. The way the different components are orchestrated has received an important inspiration from the robotics community, lately complemented with cognition-inspired approaches. In parallel, safety is a fundamental consideration in any transport-oriented AI-enabled system, where human-in-the-loop mechanisms are expected to operate in a long transition period. As a result, different approaches have emerged with the ambition to put together these different initiatives. This study aspires to briefly introduce them, highlighting their specificities, advantages, and drawbacks.

Keywords

Decision architecture; Autonomous vehicles; Cognition-inspired architecture; Embodied decision

2.1: Introduction

The decision system architecture of an automated vehicle (AV) plays a crucial role in its performance. The way different components are orchestrated has received an important inspiration from the robotics community, lately complemented with cognition-inspired approaches. In parallel, safety is a fundamental consideration in any transport-oriented AI-enabled system, where human-in-the-loop mechanisms are expected to operate in a long transition period. As a result, different approaches have emerged with the ambition to put together these different initiatives. This chapter aspires to briefly introduce them, highlighting their specificities, advantages, and drawbacks.

The outline of the chapter is as follows. The first section is devoted to provide some reflections on the influence of embodiment in the cognitive abilities that are intended to be exploited in AVs. Then, the links between cognition-inspired architectures and biological plausible human behavioral models are evoked to briefly describe the main existing decision architectures in embodied systems. The fourth section focuses to provide examples of the application on real vehicles of the previously defined families of architectures, ranging from the most extended approximation, the subsumption scheme, to shared control architectures, passing through existing ADAS and safety-oriented decision systems or specific cognition inspired architectures. A final section is focused in identifying the commonalities in most of these examples, which will be the basis of the elements described in Chapters 3–8 of the