An interpretable decision-making model for autonomous driving

Modeling the interactive behavior of human drivers is essential for achieving safe and fully autonomous vehicles. Unfortunately, most decision-making systems employed in current autonomous vehicles rely on complex deep neural network models that function as black boxes with opaque reasoning that hampers human interpretation. Drawing upon the needs theories endorsed by psychologists and driving-related psychological research, we summarize five fundamental driving needs underlying the driver’s behavior: safety, dominance, achievement, order, and relatedness. Leveraging the behavior selection module from general cognitive architectures, we propose a decision-making model explicitly tailored for autonomous vehicles, comprising three distinct modules: needs assessment, motivation generation, and behavior selection. We conducted experiments to evaluate the proposed model using a self-developed 2D simulator based on Unity. The results intuitively visualized the motivation and behavior of self-driving vehicles. This model demonstrates remarkable proficiency in handling routine tasks, such as independent and complete driving tasks, intersection navigation, and maneuvering among multiple vehicles.


Introduction
Accurate modeling of the interactive behavior of human drivers is crucial for achieving safe and full vehicle autonomy.It enables better prediction of human driver intentions and movements and is valuable for generating more human-like decisions and trajectories in autonomous vehicles.However, the current autonomous vehicle decision-making systems rely predominantly on complex deep neural network models that operate as black boxes with opaque reasoning that hinders human interpretation.][3] This paper presents a decision-making model that considers the psychology of human drivers.Drawing upon Maslow and Reiss' theories, 4,5 we consider every action driven by an underlying motive.Therefore, based on various psychological theories of needs and driving-related psychological research, this study summarizes and analyzes five fundamental driving needs that underlie driver behavior: safety, dominance, achievement, order, and relatedness.Referring to the behavior selection module of general cognitive architectures, we designed a decision-making model for autonomous vehicles based on these driving needs.The model comprises three modules: needs assessment, motivation generation, and behavior selection.The needs assessment module continuously monitors the real-time level of each need; the motivation generation module compares the priorities of urges conveyed by the needs assessment module to determine the urge with the highest priority.Finally, the behavior-selection module selects an action that aligns with the determined motivation.

Related work
Autonomous-driving decision-making has two primary approaches: behavior planning and end-to-end learning. 6Behavior planning models, particularly those based on conditional representations, such as finitestate machines, are simple and effective decisionmaking methods.They discretize the action space and are well-suited for handling the interdependent behaviors of traffic participants in natural traffic environments, where the number of interactions exponentially increases with the number of individuals.For example, MIT's Talos utilizes a navigation module for high-level behavioral planning in scenarios such as intersections, crossing, merging, and passing. 7Mellon's Boss takes a less granular approach by categorizing scenarios into three environments: roads, intersections, and zones.Correspondingly, the behavior layer of Boss distinguishes among three behaviors: lane driving, intersection handling, and achieving a zone pose. 8However, finite-state machines, which are widely employed, have limitations in coping with large and complex systems, making it challenging to scale them up in line with the complexity of driving environments.A substantial number of states and conditions must be considered to extend these methods to more general urban traffic scenarios.][11] In contrast to finite state machines that rely entirely on determined state transfer and decision rules, the partially observable Markov decision process (POMDP) provides a framework to address dynamic decision problems in imperfectly observable states. 6The Karlsruhe Institute of Technology has been exploring the application of POMDPs in decision-making for autonomous driving in the post-DARPA era.In 2014, Brechtel et al. introduced a continuous POMDP-based decision process that incorporated the poses and speeds of relevant road users.The experimental results demonstrated successful merging of the ego vehicle into the target lane without severe occlusion of the perceptual system. 12In 2017, Hubmann et al. proposed a unified POMDP decision framework that not only considered the intentions and predicted movements of surrounding vehicles, as in previous works, but also focused on intersections of any geometry with any number of traffic participants.By pre-planning routes and speeds prior to simulation, the authors reduced the action space for real-time planning. 9The exploration of POMDP at the Karlsruhe Institute of Technology has yielded promising results in simulating unsignalized intersections.However, applicable scenarios are currently limited, and the authors hope that future research will expand the scope to complex environments such as multiple lanes.In traffic environments, autonomous vehicles require frequent decision-making.However, the challenge with POMDP lies in the timeconsuming computational process and the difficulty of real-time application in online driving environments unless the complexity of the action space or environment can be reduced. 6he end-to-end approach to automated driving can be traced back to the development of the ALVINN by Carnegie Mellon University in 1989. 13Inspired by ALVINN, the NVIDIA team introduced a convolutional neural network in 2016 that directly generated steering control commands for vehicles using raw images from three front-facing cameras.This network is extensive and features 20 million connections and 250,000 parameters.The experimental results demonstrated the ability of the test vehicle to successfully navigate various road environments and diverse weather conditions. 14In 2018, Bansal et al. developed the ChauffeurNet decision-making neural network.It uses a reconstructed top-down image generated through perception system processing as the input.This image encapsulates complex urban-environment information such as vehicles, roads, signals, and a predefined route.The image is first processed by FeatureNet to extract the features, which are then used by AgentRNN to generate driving paths.Finally, the controller converts these paths into steering-wheel angles and accelerations.The ChauffeurNet model successfully operates a realworld vehicle although improvements can be made when compared with traditional motion-planning methods. 15Chen et al. adopted a similar technical approach at the University of California, Berkeley.They developed a deep imitation learning trajectory planning module utilizing a convolutional neural network with a fully connected layer containing 1000 units connected after the VGGNet16.The output layer represents the relative coordinates of the target sequence.Despite encountering errors in roundabouts, unstructured areas, and two-lane roads, the model demonstrated the ability to follow designated routes in multiple road environments.It responded to surrounding objects and adjusted to different traffic signal states, achieving results using only 120,000 data frames. 16ome studies argue that the behavior planner, tasked with making high-level decisions, and the trajectory planner, responsible for generating feasible trajectories, are commonly independently developed, lacking a shared objective.This separation changes the behavior planner, adversely affecting the fine-tuned trajectory planner, thus failing to fully exploit their strong coupling.Consequently, studies have attempted to address this issue by integrating behavior planning with motion planning.Uber's Sadat et al. proposed an approach featuring an interpretable cost function and a joint learning algorithm that learns a shared cost function employed by both their behavior and trajectory components.Experiments have demonstrated that the jointly learned planner performs better in terms of both similarity to human driving and other safety metrics. 17n another study, Singh et al. proposed a bi-level optimization that can simultaneously search for the optimal higher-level behavioral decisions along with the lower-level trajectories necessary for executing them.Extensive simulations demonstrated that their approach outperforms state-of-the-art model predictive control and reinforcement learning approaches in terms of collision rate while remaining competitive in driving efficiency. 18lthough safety and ethical concerns persist, 19 both the end-to-end approach and other AI-based decisionmaking methods have achieved significant advancements.Current automated driving decision methods rely predominantly on deep neural networks and fall under the connectionist approach of cognitive science.These models often process data in a black-box manner and lack interpretability.Although these models yield accurate prediction results, explaining their underlying logic and decision-making processes has become challenging.In the event of problems during simulation tests, the interpretability of autonomous vehicle decisions is crucial for analyzing the triggering factors of accidents and promptly troubleshooting the tested autonomous driving system. 20,21cision-making model

General structure
The decision-making model presented in this paper was inspired by the CLARION 22 and MicroPsi 23 cognitive architectures.Figure 1 illustrates this model following a driving-need-driven approach.It comprises three modules: needs assessment, motivation generation, and behavior selection.The needs assessment module continuously monitors real-time needs.When an autonomous vehicle experiences an urge for one or more needs in its current environment, the needs assessment module sends the urge to the motivation-generation module.The motivation-generation module compares the priorities of the urges received from the needs-assessment module.Urges with higher levels are assigned higher priority and become the driving force.Even if an action selected at a lower level is already in progress, it is immediately terminated to relinquish decision dominance.
Finally, the behavior selection module makes further assessments and determines the appropriate behavioral action that aligns with motivation, including the goal of the action and the timeframe to achieve it.

Needs assessment
Real-time evaluation and monitoring of the urge u d i for each need d i were conducted using the needs assessment module.When an autonomous vehicle experiences urges for one or more needs in a given environment, the needs assessment module promptly transmits these urges to the motivation generation module.Table 1 illustrates the five driving requirements: safety d sfty , dominance d dom , achievement d achv , order d ord , and relatedness d rel .These needs are derived from various human needs theories and through extensive research on the psychological motivations of drivers.The corresponding urges for each requirement are described below.
Urge for safety.Safety is a fundamental requirement for survival, which compels individuals to avoid risks and potential harm when faced with external threats.The significance of safety needs in driving is evident from the fact that, as of 2008, approximately 10% of the articles published in the journal Human Factors focused on driving and driving safety. 33Meeting safety needs is crucial for drivers because it allows them to feel secure and comfortable while driving, minimizing excessive psychological loads.Drivers are compelled to maintain  an adequate distance from stationary entities, such as road boundaries and obstacles, and dynamic entities, such as vehicles and pedestrians.Summala's ''multiple comfort zone'' model introduces the concept of a ''safety margin,'' which defines safety requirements in terms of time and space.Once an entity enters the driver's subjectively constructed ''safety margin,'' the driver experiences discomfort and fear, prompting them to take appropriate actions, such as braking or steering, to pull away and mitigate potential risks. 34,35he urge for safety u sfty was exclusively determined by the safety zone zone sfty , which is a virtual rectangular area surrounding an autonomous vehicle.This region represents the minimum safe distance that an autonomous vehicle aims to maintain from other dynamic entities.Its dimensions are influenced by the size, speed, and acceleration of the vehicle.When any other dynamic entity entity dyn enters this region, the safety need d sfty triggers a safety urge u sfty , as indicated by equation (1).
Urge for dominance.Dominance refers to the desire to exert influence over the environment and fulfill one's aspirations.In the context of driving, the need for dominance is evident through a driver's active control of speed and direction.Summala suggests that whether driving at a normal speed or pushing the limits for excitement, drivers experience a sense of ''driving pleasure,'' which arises from fulfilling their desire for control over the vehicle. 36The discretionary lane changes observed in the microscopic traffic-flow models can be attributed to the need for dominance. 37Drivers can change lanes to adjacent lanes to maintain their desired speed without being impeded by slower traffic in the same lane, thereby ensuring control over the driving situation.
The need for dominance d dom encompasses the driver's desire to control the vehicle's direction and speed, particularly in the pursuit of higher speeds.It generates an urge u dom in two specific situations: when there is another dynamic entity entity dyn within a certain distance ahead, and when there is a discrepancy between the current v curr and desired speeds v expt .
where u dom dyn is the impulse related to the front vehicle during the control urge as expressed in equation 3 ð Þ and u dom v is the urge related to vehicle speed, as shown in equation 4ð Þ.
Urge for achievement.This requirement for achievement is typically purposeful.Accomplishing a goal or task is necessary to achieve ambition. 38Summala's ''good or expected driving progress'' concept captures the driver's need for achievement. 34In microscopic traffic models, mandatory lane changing is typical behavior driven by the need for achievement, particularly when reaching a destination. 35However, Vechione et al. present slightly biased situations to justify mandatory lane changes, such as unsustainability in the current lane or the need to cooperate with intersection turns. 39onetheless, these situations only partially encompass the range of mandatory lane changes.For example, if a driver intends to travel straight and the left-turn lane is not part of the planned route, a mandatory lane change is necessary if the driver is in the left-turn lane at that point.Therefore, a mandatory lane change occurs when the current lane deviates from the planned route.Conversely, when drivers are already in a lane that aligns with the planned route, they would refrain from making a lane change to fulfill their driving task, driven by the need for achievement.When establishing a planned route from the initial point to the destination, it is imperative that autonomous vehicles adhere to this route.Hence, throughout the journey, an autonomous vehicle must continuously assess whether the current lane aligns with the planned route.This evaluation involves determining whether a lane change is required.Specifically, if the estimated distance for a lane change dst curr!dest exceeds the feasible lane change distance dst 0 rmng , the need for achievement d achv will generate an urge for achievement u achv .Conversely, if the estimated distance for a lane change dst curr!dest falls short of the executable lane-change distance dst 0 rmng , the need for achievement d achv will not generate the urge for achievement u achv , as illustrated in equation (5).
Before determining the achievement urge, the nearest candidate lane lane dest that aligns with the predetermined route Route to the current lane lane curr must be identified.Assuming that several lanes are available within the current section of the road that lead to the preplanned route, the set comprising these lanes can be represented by equation (6).
The lanes nearest the current lane from the set CandidateLaneSet are selected, and the function ID x ð Þ is used to obtain the lane number. where Because more than one candidate target lane closest to the current lane lane curr can exist, all candidate target lanes are represented as a set TargetLaneSet, and one lane is randomly selected as the target lane lane dest .
Next, the estimated lane-change distance dst curr!dest from the current lane to the target lane is calculated.Assume the following: Using the average speed v ( i of each lane lane i in section Sect curr , we estimate the distance to travel from the current lane lane curr to the target lane lane dest in the longitudinal direction dist curr!dest . where t w is the estimated lane-change waiting time and t c is the estimated lane-change duration.The executable lane change distance dst 0 rmng can be calculated using equation (11).
where dist 0 is the reserved distance of the road section and dist rmng is the actual remaining distance from the current position to the end of the road.
Urge for order.The need for an order pertains to adhering to regulations and avoiding penalties.In the context of driving, these regulations include formal traffic laws, 36,40 such as obeying traffic signals, staying within designated lanes, and adhering to speed limits, and informal rules. 41Adhering to these rules helps drivers avoid accidents and penalties, 33 enhances the overall efficiency driving, minimizes congestion and accidents, and saves time and expenses for both drivers and other road users.A straightforward example is when approaching a signalized intersection that follows the instructions displayed by traffic signals, either coming to a stop or proceeding, depending on the signal's indication.In the absence of traffic signals or police instructions, drivers follow rules applicable to unsignalized intersections, such as yielding to straight-through traffic, giving way to vehicles on the main road at a junction, and prioritizing right turns.While the need for order is relevant throughout the journey, certain actions, such as staying within lanes and adhering to traffic regulations, are already implicit in the motion planning process.Therefore, the current model explicitly incorporates the need for order as autonomous vehicles approach intersections.As an autonomous vehicle approaches an intersection, the need for an order d ord triggers the order urge u ord , as described in equation (12).
An order urge is generated when the distance dst x between the vehicle and intersection falls below the desired distance dst rsv .Conversely, no order urge is generated when the distance from the intersection exceeds the desired distance.
Urge for relatedness.The need for relatedness encompasses the desire to establish connections with others, manifesting as positive experiences, such as acceptance, respect, and reciprocity, or negative experiences, such as refusal to cooperate, rejection, and confrontation.While Summala's ''multiple comfort zones'' model primarily focuses on individual needs, it does not extensively address interpersonal dynamics. 33,34Rumar defines relatedness as a traffic objective that involves interacting with other traffic participants in a manner that maintains mobility while avoiding collisions. 36owever, achieving this goal in practice is challenging.
For instance, when a vehicle in an adjacent lane merges into the driver's lane, the driver may adjust the speed and increase the following distance to facilitate merging if they accept or agree with the other vehicle's behavior.Conversely, if the driver rejects or opposes merging, they may maintain or slightly increase their speed while reducing the following distance to prevent the other vehicle from merging.It is evident that maneuverability is compromised in the former scenario, whereas collision risk is heightened rather than mitigated in the latter scenario.
The need for relatedness encompasses the desire to cooperate with other traffic participants, seek approval and recognition, and express refusal.In the context of autonomous vehicles, other dynamic entities entity dyn actively communicate their requests to the vehicle when affected by its actions.An autonomous vehicle generates a relatedness urge upon receiving a request REQ dyn , as expressed in equation (13).

Motivation generation
Psi's theory proposes that all goal-directed actions stem from motivation. 42Reproducing these motivations enhances the comprehensiveness of the decision model and provides a deeper understanding of the underlying mental processes.Once the environment triggers the generation of urges corresponding to needs, the motivation generation module compares the priorities of all available urges.The urge for higher priority is the prevailing motivation.Even if an action planned at a lower level is currently being executed, it must be terminated immediately to relinquish the dominance of the decision.The priority order of each urge within the motivation generation module is outlined below.p sfty .pord .pachv .prel .pdom ð14Þ In the above equation, p i denotes the priority of the urge u i .Safety urge had the highest priority, followed by order, achievement, relatedness, and dominance.In the absence of an urge, the decision-making process enters an unmotivated behavioral phase.

Behavior selection
Alternative behaviors and actions.This section confines the actions that the driver can perform B to changing the speed B changeSpeed , changing the lane B changeLane , and maintaining the current state B unchange , as shown in equation (15).
Changing speed.The changing speed B changeSpeed refers to the behavior of adjusting the velocity of the vehicle, either by increasing or decreasing the speed to reach a specific target.This behavior includes maintaining distance a cs keep dst , emergency braking a cs emg brk , adjusting speed a cs adj lmt , slowly stopping a cs stop slw , and yielding a cs plt dec , as shown in equation (16).
B changeSpeed = a cs keep dst , a cs emg brk , a cs adj lmt , a cs stop slw , a cs plt dec Both action a cs keep dst and action a plt dec involve maintaining a specific distance from the preceding vehicle; however, they differ in their target objects.a cs keep dst aims to maintain a safe distance from the vehicle directly ahead in the current lane, whereas a plt dec targets the vehicle ahead that is approaching to merge into the current lane.The completion time t f of the two actions is the headway time distance t pre h of the driver preference.The target states are expressed using equation (17), where x f , v f , and a f denote the target position, target speed, and target acceleration at the end of the maneuver, respectively; x veh fnt , v veh fnt , and a veh fnt are the position, speed, and acceleration of the target vehicle, respectively.
a cs emg brk and a cs stop slw aim to bring the vehicle to a stop with a target speed of zero.However, they differ in execution characteristics.a cs emg brk focuses on stopping the vehicle as quickly as possible, resulting in shorter execution time without specifying the target location.By contrast, a cs stop slw has a predetermined target location for the vehicle to stop.
Finally, a cs adj lmt is designed to enable the vehicle to reach the target speed within a specific timeframe.The target speed in this action is determined by the driver's desired speed v exp considering the road conditions.The execution time is influenced by driver comfort preferences and vehicle performance.
Changing lane.Changing lane B changeLane refers to the transition from one lane to another.This behavior includes four actions: mandatory lane changing a cl dest ln , discretionary lane changing a cl obs ln , random lane changing a cl rnd ln , and aborting lane changing a cl abt ln , as shown in equation (18).
Mandatory lane changing.a cl dest ln involves changing lanes toward the nearest target lane.While multiple lanes may separate the current lane lane curr and target lane, this action only executes a single lane change, as illustrated in Figure 2. In contrast to the random lane selection utilized to generate the achievement urge for distance estimation, the target lane for mandatory lane changing and the lane for a cl dest ln action can be determined through the following processes: After determining the set TargetLaneSet, the set of lanes for a single lane change TempLaneSet is the lane that is one unit away from the current lane lane curr in the direction of the mandatory target lane. where Li et al.
We then check if the lanes in TempLaneSet have sufficient space to support a lane change.
where spareSpace Á ð Þ is a function that determines whether space exists in a lane.If more than one lane still exists in the set AvailableLaneSet, then any lane can be considered as the single lane-changing target lane t of action a cl dest ln , and lane dest is the corresponding target lane for mandatory lane changing.

Discretionary lane changing. a cl
obs ln changes lanes to adjacent lanes, whose average speed is faster than that of the current lane.The set of adjacent lanes AdjacentLaneSet is determined using equation (22).
We then filter adjacent lanes using average speed to retain lanes with faster average speeds.
where denotes the current lane average speed and v ( lane denotes the lane lane average speed.We check whether the lanes in FastLaneSet have sufficient space to support lane changes.
If set AvailableLaneSet still has two lanes, we select any lane from it as the target lane for this action.
Random lane changing.a cl rnd ln randomly selects adjacent lanes for lane changes.After the set of adjacent lanes, AdjacentLaneSet is determined using equation ( 22), the lanes are identified to determine whether there is sufficient space to support lane changes.
If there are still two lanes in set AvailableLaneSet, any lane from it is taken as the target lane for this action.
Aborting lane changing.During the execution of any of the aforementioned lane-changing actions, a cl abt ln can interrupt the ongoing lane-changing and revert to the lane before initiating the lane-changing action.
Maintaining unchanged.B unchanage involves neither lane changing nor altering the speed of the vehicle, and instead maintains its current speed v curr and continues along the current lane lane curr .This set of behaviors encompasses both the normal maintenance action, a cn mnt unch , and the alert-and-maintain action, a cn alert .It is essential to highlight that a cn alert emits a cooperative signal to other traffic participants who will respond to the signal based on their own needs and preferences.
Behavioral decision for each motivation.In this study, the behavioral decisions for each motivation were represented using individual behavior trees.Behavior trees are widely adopted for designing and implementing intelligent behaviors in various systems.They offer a hierarchical structure that enables the creation of complex and adaptive behaviors by organizing smaller behavioral units known as nodes.
The execution of a behavior tree commences at the root, which sends ticks to its child nodes at a predefined frequency.A tick serves as an enabling signal to execute a child node.When a node in the behavior tree is eligible for execution, it returns a status to its parent, indicating whether it is ''running'' if its execution is ongoing, ''success'' if it has achieved its goal, or ''failure'' if it has not.
This study utilized three control nodes (sequence, selector, and random selector), two execution nodes (action and condition), and one decorator node (inverter).The sequence node (Figure 3  Actions for safety motivation.The behavioral decisions driven by safety motivation are shown in Figure 4.It begins by determining whether the hazard originates in the longitudinal or the lateral direction.Suppose that the dynamic entity ahead entity dyn approaches the autonomous vehicle's safety zone too closely, triggering the safety motive m sfty , the ego vehicle initiates the action a cs keep dst to adjust the distance between the two vehicles.If this action fails to eliminate the imminent danger, and entity dyn remains within the safety zone zone sfty , a cs emg brk is activated to rapidly reduce the speed of the vehicle, with the aim of creating a more significant gap between the vehicles until the hazard is resolved or the vehicle stops.
An additional emergency braking action, a cs emg brk , is included in the sequence node to account for potential motion planning failures in a cs keep dst .A selector connects the two parts.
Suppose that the ego vehicle is in the process of changing lanes and becomes too close to a lateral adjacent dynamic entity.In this case, the lane-change action must be promptly aborted and the vehicle should be returned to the original lane via a cl abt ln .If the autonomous vehicle is not engaged in a lane-change process, it decelerates using a cs emg brk and swiftly moves away from the longitudinal direction to avoid a collision with the adjacent lateral entity.
Actions for dominance motivation.When a dominance motive is present, it is essential to differentiate whether it arises from an urge associated with the preceding vehicle u ctl dyn or from an urge related to the speed of the ego vehicle u ctl v .If the preceding vehicle enters the ego vehicle's field of view and its speed is slower than the ego vehicle's current speed, the motive is related to the preceding vehicle.In this case, the ego vehicle can maintain its current motion and signal the preceding vehicle via a cn alert , to either appropriately increase its speed or change lanes to faster adjacent lanes.
If the current lane is unobstructed by the preceding vehicle, or if there is no preceding vehicle at all and the ego vehicle's speed is lower than the desired speed, it suggests that the dominance motive is speed-dependent.In this scenario, the ego vehicle can engage in a speed adjustment action a cs adj lmt .The behavioral decisions driven by dominance motivation are shown in Figure 5.
Actions for achievement motivation.The behavioral decision process driven by the achievement motive begins by assessing whether the current lane lane curr is among the set CandidateLaneSet.If it is not considered a candidate target lane, a cl dest ln is required to transition toward lane dest .Conversely, if the current lane corresponds to the target lane, the driving state remains unchanged via a cn mnt unch because the remaining distance is deemed insufficient at this stage.The behavioral decisions driven by achievement motivation are shown in Figure 6.
Actions for order motivation.Behavioral decisions driven by the order motive are currently governed by traffic rules applicable to unsignalized intersections that dictate that vehicles proceed sequentially based on their arrival order at the intersection.If there are preceding vehicles traversing the intersection, vehicles approaching the intersection must stop at the designated stop line via a cs stop slw and must maintain their stationary state until it is their turn to proceed via a cn mnt unch .The behavioral decisions driven by order motivation are depicted in Figure 7.
Actions for relatedness motivation.The behavioral decision process driven by the relatedness motive is divided into three distinct scenarios.First, suppose that the ego vehicle is undergoing a lane change when it receives a  request from the vehicle behind it.In this case, it has the option to either proceed with the ongoing lanechange action or abort the lane change and return to the original lane via a cl abt ln , based on a predefined preference probability.Given the ambiguity of the source of the lane-change action at this point, a proxy action a cl delegate is employed to represent the ongoing lanechange action originating from a lower-level motive.
Second, suppose that the ego vehicle is in a normal driving state and does not undergo a lane change when it receives a request.In this case, it can respond by selecting one of three actions: adjusting the speed a cs adj lmt , maintaining the current motion a cn mnt unch , or randomly changing lanes to an adjacent lane a cl rnd ln .Third, when the ego vehicle receives a request from the vehicle ahead, a possible scenario arises in which a vehicle in an adjacent lane intends to merge into the ego vehicle's lane.In response, the ego vehicle can choose to either accept the request and ensure a safe distance by slowing down and yielding via a cs plt dec or can decline the request and maintain its current state to prevent lane change via a cn mnt unch .The behavioral decisions driven by the relatedness motivation are shown in Figure 8.
Unmotivated.When none of the motives are fulfilled, the autonomous vehicle resorts to selecting one of the two actions: a cn mnt unch and a cl rnd ln .To mitigate the risk of motion planning failure associated with action a cl rnd ln , a cn mnt unch was incorporated outside the random selector node.This inclusion ensures reliable functioning of the behavioral decision process.The behavioral decisions of the unmotivated state are shown in Figure 9.

Experiments
The autonomous driving model employed for the experimental validation in this study adopts Michon's three-level driver behavior model, which includes strategic, maneuvering, and control levels. 43At the strategic level, path planning was accomplished using the A* algorithm based on the road network.The maneuvering level was further divided into behavioral and motion planning.The decision-making process utilizes the model proposed in this study, whereas the motion planning aspect employs the analytical calculation method outlined by Li. 44 Lastly, the control level assumes an ideal ''unit 1'' scenario, where vehicle control and dynamics are optimized.In other words, the ego vehicle precisely follows the trajectory generated by the planning module.
The experiments were performed on a MacBook Pro (16-inch, 2019) with a 2.3 GHz Intel Core i9 processor, 32 GB 2667 MHz DDR4 memory, and AMD Radeon Pro 5500M with 8 GB of graphics memory.The software employed in this study was a custom-developed 2D simulator built on Unity 2022.2.0b12, 45 and all programming tasks were implemented using the C# language.Each vehicle within the simulation utilized the same decision-making model, but with distinct preference settings.Given the multifaceted nature of the model, the evaluation process involved conducting three separate experiments to validate the different features.The experiments were conducted as follows: 1. Completion of the assigned driving task: In this experiment, a vehicle traveled from its designated starting point to a predefined destination.The objective was to assess the ability of the model to navigate predetermined routes effectively.
2. Handling unsignalized traffic junctions: Four vehicles converge at a junction sequentially, and each vehicle must wait for the preceding vehicles to exit the junction before entering.This experiment focused on the capability of the model to navigate through unsignalized junctions.3. Weaving ability in multilane traffic: A vehicle is expected to maneuver skillfully through groups of vehicles from the rear of a multilane highway, allowing it to increase its speed.This experiment evaluated the weaving ability of the model in multilane traffic situations.

Completion of assigned driving task
The autonomous vehicle embarked on its journey from lane 1 on Road 2392, with the ultimate destination being Road 2452.The planned route, indicated by the green lines in Figure 10 Beginning at t = 0 s, the vehicle operates under dominant motivation, selecting action a cs adj lmt to increase its speed to approximately 10:8 m/s.Subsequently, it transitioned to an unmotivated state and opted for the action a cn mnt unch .At t = 86:94 s, as the vehicle reached the end of Road 2392, it entered the order motivation and decelerated to a speed of 0 m/s using action a cs stop slw .Given the absence of other vehicles at the intersection of Road 2391, the vehicle  returned to dominance motivation, initiating acceleration through action a cs adj lmt .There was an approximately 0:4 s during which the vehicle entered the achievement motivation.However, upon entering Road 2431, it reverted to an unmotivated state and continued driving via action a cn mnt unch .Because no other vehicles were involved in the scenario and only one autonomous vehicle was present, this cycle was repeated until the vehicle reached its destination, Road 2452.The entire journey lasted for approximately 11.5 min.In Figure 11, from top to bottom, the five panels represent the road ID, longitudinal displacement, motivation, action, and longitudinal velocity versus time of the autonomous vehicle, respectively.veh 1003 departed from Road 5165, turned right at the intersection, and continued on Road 5154.

Handling unsignalized traffic junctions
Each of the four vehicles began their journey at 4 s intervals.Upon starting, each vehicle entered the dominance motivation, increasing its speed to approximately 10m=s using the action a cs adj lmt .Subsequently, it transitioned into an unmotivated state and performed the action a cn mnt unch .At t = 20:32s, veh 1002 entered order motivation and initiated deceleration using a cs stop s lw .Because no other vehicles were present at the intersection, veh 1002 transitioned into dominance motivation and accelerated using a cs adj lmt .Concurrently, veh 1001 approached the intersection and entered the order motivation, starting to decelerate with a cs stop slw .At t = 28:98 s, veh 1002 exited the intersection onto Road 5165, allowing veh 1001 to transition into dominance motivation, increase its speed, and enter the intersection.At t = 28:52 s, veh 1000 entered the order motivation and began to slow down, reaching a complete stop at t = 33:08 s, awaiting the departure of veh 1001 from the intersection, as illustrated in Figure 13 a ð Þ and b ð Þ.Subsequently, at t = 33:1 s, veh 1000 commenced accelerating and entered the intersection.After a delay of approximately 2:5 s, veh 1003 arrived at the intersection.It entered the order motivation and initiated deceleration using a cs stop slw , whereas veh 1000 proceeded straight through the intersection.
By t = 39:52 s, veh 1000 transitioned to dominance motivation, accelerated using a cs adj lmt , and turned right onto Road 5307.At this point, the remaining vehicles entered an unmotivated state and continued their motion using a cn mnt unch .
From top to bottom, the five panels in Figure 14 show the road ID, longitudinal displacement, motivation, action, and longitudinal velocity versus time for the four vehicles, respectively.The blue dash-dotdotted, orange dashed, green dotted, and red dashdotted lines represent veh 1000 , veh 1001 , veh 1002 , and veh 1003 , respectively.

Weaving in multilane traffic
From top to bottom, the seven panels in Figure 15 show the longitudinal displacement, motivation, action, longitudinal velocity, lane ID, lateral displacement, and lateral velocity versus time for the six vehicles, respectively.The blue dash-dot-dotted, orange dashed, green dotted, red dash-dotted, purple dense dash-dotted, and pink solid lines represent veh 1000 , veh 1001 , veh 1002 , veh 1003 , veh 1004 , and veh 1005 , respectively.On the multi-lane highway (road 1030), vehicles veh 1000 , veh 1001 , veh 1002 , veh 1003 , and veh 1004 started from different lanes every 6 s.During this period, they were in the dominance motivation and accelerated states using a cs adj lmt .Once the vehicles reached a speed of approximately 10 m=s, they transitioned to an unmotivated state and continued their motion by executing action a cn mnt unch .At t = 30:04 s, veh 1005 departed from the À3 lane and entered dominance motivation, accelerating through a cs adj lmt .At t = 33:52 s, veh 1005 transitioned to an unmotivated state, maintaining its motion for only 0:18 s before noticing veh 1004 ahead in the same lane.It then reverted to the dominance motivation and changed lanes from the À3 lane to the À2 lane using the action a cl obs ln .After adjusting its speed with a cs adj lmt , veh 1005 entered an unmotivated state and continued its movement.
At t = 46:4 s, veh 1005 noticed veh 1003 ahead and entered dominance motivation, expecting to alert the other vehicle to increase its speed through action a cn alert .In response, veh 1003 entered the relatedness motivation and increased its speed from 9:68 m=s to 9:96 m=s using a cs adj l mt .However, the speed increase was not sufficiently significant, prompting veh 1005 to decide to change lanes again.Upon detecting that veh 1005 intended to switch to its lane, veh 1004 entered relatedness motivation and executed the yielding action a cs plt dec to decrease its speed and maintain a safe distance.At t = 53:2 s, veh 1005 successfully changed lanes back to the À3 lane, followed by veh 1004 , which resumed its speed to 10:5 m=s and returned to the unmotivated state.
At t = 65:52 s, veh 1005 caught up with veh 1001 ahead and, in dominance motivation, attempted to alert veh 1001 to increase its speed through action a cn alert .veh 1001 immediately entered relatedness motivation and increased its speed using a cs adj lmt , although the increase in speed was not significant.veh 1005 then utilized the action a cl obs ln to change lanes from the À3 lane back to

Conclusion
We summarize and analyze the five fundamental driving needs underlying driver behavior, namely, safety, dominance, achievement, order, and relatedness, based on various psychological needs theories and research on driving psychology.Subsequently, an interpretable autonomous driving decision-making model is proposed.The model consists of three modules: needs assessment, motivation generation, and behavior selection.The needs assessment module continuously monitors the levels of various needs in real-time.The motivation generation module compares the priority of urges sent by the needs assessment module and determines the urge with the highest priority as the motivation.Finally, the action behavior selection module determines the actions that align with motivation.The experiments were conducted using a self-developed 2D simulator based on the Unity software.All vehicles in the simulation were equipped with the same decisionmaking model, but with different preferences.The results provide intuitive observations of the motives and behaviors of autonomous vehicles and demonstrate their effectiveness in various driving tasks, including single-lane driving, intersections, and multi-vehicle interactions.
However, the existing approach for behavior selection is based on a behavior tree, which is a relatively rigid planning method.Once the structure is established, the method of behavior selection remains unchanged.Future research endeavors may explore alternative methods that offer greater flexibility, introducing a higher level of uncertainty in behavior selection.

Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.

Figure 1 .
Figure 1.General structure of the decision-making model.It comprises three modules: needs assessment, motivation generation, and behavior selection.
(a)) seeks and executes the first child who has not yet succeeded.The selector node (Figure 3(b)) identifies and executes the

Figure 2 .
Figure 2. Interrelationships between lanes.Current lane lane curr , target lane for mandatory lane-changing lane dest and lane for the a cl dest ln action lane t .

Figure 3 .
Figure 3. Nodes in the behavior tree.This study utilizes three control nodes ((a) sequence, (b) selector, and (c) random selector), two execution nodes ((d) condition and (e) action), and one decorator node ((f) inverter).

Figure 4 .
Figure 4. Behavior tree for safety motivation.

Figure 7 .
Figure 7. Behavior tree for order motivation.

Figure 9 .
Figure 9. Behavior tree for the unmotivated state.

Four
autonomous vehicles, namely veh 1000 , veh 1001 , veh 1002 , and veh 1003 , approached the intersection on different roads, as depicted in Figure 12.The routes followed by each vehicle are as follows: veh 1000 departed from Road 5277, proceeded straight through the intersection, and entered Road 5154; veh 1001 departed from Road 5271, turned right at the intersection, and entered Road 5277; veh 1002 departed from Road 5154, turned left at the intersection, and entered Road 5165; and

Figure 11 .
Figure 11.Completion of the assigned driving task.From top to bottom, the five panels (a-e) represent the road ID, longitudinal displacement, motivation, action, and longitudinal velocity versus time of the autonomous vehicle, respectively.

Figure 12 .
Figure 12.Routes followed by each vehicle.veh 1000 departed from Road 5277, proceeded straight through the intersection, entered Road 5154, veh 1001 departed from Road 5271, turned right at the intersection, and entered Road 5277, veh 1002 departed from Road 5154, turned left at the intersection, and entered Road 5165; and veh 1003 departed from Road 5165, turned right at the intersection, and continued onto Road 5154.

Figure 14 .
Figure 14.Handling unsignalized traffic junctions.From top to bottom, the five panels (a-e) represent road ID, longitudinal displacement, motivation, action, and longitudinal velocity versus time for the four vehicles.The blue dash-dot-dotted, orange dashed, green dotted, and red dash-dotted lines represent veh 1000 , veh 1001 , veh 1002 and veh 1003 , respectively.

Figure 15 .
Figure 15.Weaving in multilane traffic.From top to bottom, the seven panels (a-g) respectively represent the longitudinal displacement, motivation, action, longitudinal velocity, lane ID, lateral displacement, and lateral velocity versus time for six vehicles.The blue dash-dot-dotted, orange dashed, green dotted, red dash-dotted, purple dense dash-dotted, and pink solid lines represent veh 1000 , veh 1001 , veh 1002 , veh 1003 , veh 1004 , and veh 1005 , respectively.

Table 1 .
Correspondence of the five driving needs in different theories.24-32