【Paper Notes】A Review of Motion Planning for Highway Autonomous Driving

I. INTRODUCTION

First, it introduces the origins of some self-driving cars, the application of related assisted driving technologies, related policies and standards, the DARPA Challenge and other self-driving car project situations and progress.

---

The author introduces some previous reviews and literature on Motion Planning:

• Robotics field: [20], [21], [22];
• Autonomous driving: [23], [24], [25];

---

• The standardization of research to adapt to industrially realistic prototypes of autonomous vehicles brings new issues such as functional safety, real-time computing, systems approaches or low-cost development, as well as application to specific scenarios in parking lots, intersection management or highway driving. .
• Reasons for studying autonomous driving on highways :
  • The environment is simple and the road structure is simple;
  • Driving behavior is limited and unsophisticated;
  • Easier to achieve fully autonomous driving.

II. CONSIDERATIONS FOR HIGHWAY MOTION PLANNING

A. Terminology

This part introduces some basic terms and concepts (not fully documented):

• configuration space : geometric state space (robot configuration space).
• Evolution space : refers to the area where the vehicle can move safely in space and time.

Both include:

• collision space : Will collide with obstacles;
• Uncertain space : There is a probability of collision with obstacles;
• free space : Will not collide with obstacles.

---

• The author distinguishes the generation of constructing paths, trajectories, maneuvers or action sequences from planning.

B. Motion Planning Scheme

• [18], [24], [27] These three documents describe the hierarchical architecture of autonomous driving.

The author introduces a typical layering pattern:

• Motion strategy input: related information about self-vehicle, obstacle behavior and infrastructure obtained from perception , localization and communication
  ( Vehicle-to-X-V2X ) modules. (Not focused on how to collect, but focused on the quality, uncertainty and reliability of measurements )
• Scene representation : Provides a perceptual map including the vehicle, obstacles, surrounding environment, traffic, etc.
• From the perception, localization and communication (Vehicle-to-X -
  V2X) modules to the motion strategy module, a closed-loop system is formed: transmitting some current and future information.
• control: control, execution (actuation).
  Closed-loop system: Feeds back some status of the current vehicle to the Motion part.
• Interaction with the driver: warns the driver of risks; feeds back the driver’s intentions.

---

The layering approach proposed by the author in this article:

• (i) route planning : global planner;
• (ii) prediction : prediction, inferring changes in the surrounding environment based on current and historical information;
• (iii) decision making
• (iv) generation
• (v) deformation

Deformation refers to the process of fine-tuning the generated motion trajectory in motion planning to adapt to the environment and specific constraints , thereby improving the accuracy and feasibility of motion planning . This process can be divided into two stages: high-level strategy and low-level strategy . In the high-level strategy stage, multiple possible motion trajectories are generated and the most appropriate trajectory is selected based on specific evaluation criteria; in the low-level strategy stage, adaptation to the environment and constraints is achieved by fine-tuning the trajectory. Deformation can also be viewed as a reactive strategy that can adjust in real time on the generated motion trajectory to avoid obstacles and maintain trajectory continuity .

---

In recent work [32], [33], the motion planning method is mainly divided into two parts: high-level and low-level:

• High-level (predictive): Contains three parts: risk evaluation, criteria minimization, and constraint submission. These three parts are mainly used in the decision-making part: selecting one from a series of planning trajectories as a decision; or making a decision first and then planning accordingly. High-level benefits from longer field of view prediction, but is more time-consuming.
• low-level (reactive): Replace the motion generation process with a reaction-based method.

The table shows the distance range and time range of each module.

C. Specificities of Highway Driving

The scenarios that the author mainly considers are limited to one-way traffic flows on structured roads with speeds exceeding 60km/h. Compared with urban road scenes, traffic behavior is simpler and more uniform.

The author mainly focuses on eight situations in this article:

• Lane keeping；
• Car following；
• Lane changing；
• Lateral-most lane changing : changing lanes to the outermost lane;
• Passing；
• Overtaking；
• Merging；
• Highway toll : high-speed toll station.

---

• Motion planning for formation driving requires stronger robustness and stability. Related literature: [35].
• The main difference between high-speed motion planning and urban road-based motion planning is the preview time . The former has a longer preview distance/time.

D. Constraints on Highway Driving

• Hard constraints : considering traffic rules and avoiding collisions
• Soft constraints : consider time, distance, energy consumption minimization, and comfort.
• Other feasibility constraints : vehicle non-integrity constraints; smooth path - the trajectory needs to be differentiable and continuous curvature; vehicle dynamics constraints.
• Except for methods based on virtual potential fields and numerical optimization methods, most methods have little or no consideration of vehicle models .
• Time-varying dynamic constraints (e.g. monotonicity?).

E. What Is at Stake in this Paper

This section mainly focuses on the motion strategy algorithm of highway autonomous driving, emphasizing real-time effectiveness and not sticking to mathematical complexity (Simple is the BEST!). The authors classify algorithms from a practical application perspective and analyze their advantages and disadvantages, taking into account simulations or experiments. Compared with optimality, feasibility is more important. In highway driving, the key is to ensure at least one continuous motion strategy.

III. STATE OF THE ART

Motion Planning related questions:

1. The data passed by the scene representation will be of different types (discrete/continuous, algebraic/analytical or static/dynamic), and the form of the motion planning problem will be different accordingly.
2. Motion planning combines five inevitable aspects, (i) state estimation, (ii) time evolution, (iii) action planning, (iv) standard optimization, and (v) compliance with constraints. How these are dealt with changes the corresponding issues.

A. Taxonomy Description

Describe how the author classifies the methods of Motion Planning

Classification features:

• Type of output: space, path, trajectory, operation or symbolic representation.
  • At the same time, it is also necessary to consider whether the algorithm returns decomposition motion or reference motion. The former is called set-algorithm and requires a complementary algorithm to find feasible motion; the latter is called solve-algorithm ;
• Space-time properties: predictive or reactive horizon;
• Mathematical domains (methods) used: based on geometry, heuristic, logic, cognitive, biomimetic ( geometric, heuristic, logic, cognitive, biomimetic ).
  • The geometric method is based on the properties of space and directly handles the spatial constraints (kinematic constraints) of the environment and its own vehicle. Next, we need to deal with large space exploration and optimization problems.
  • The heuristic heuristic method relies on knowledge about the problem to quickly find approximate solutions and avoid computational complexity and road congestion, but it does not guarantee that the optimal solution will be found.
  • Logic logic method refers to the deductive method based on inference, which is easy to connect results and causes, but is susceptible to combinatorial explosion.
  • Cognitive cognitive methods rely on prior knowledge and common sense to evaluate situations and imitate the human decision-making process. However, there is insufficient experience in autonomous driving applications, the theoretical validity has not been verified, and cognitive-based algorithms are not recognized. Driver behavior is complex and irrational, making it difficult to model.
  • The biomimetic approach, which describes physics-inspired algorithms that follow physical laws that are easy to implement but can lead to infinite loop motion behavior. Feasible solutions must be used to obtain convergence of the system.

Classification radar charts of several algorithms to be introduced next

B. Algorithm Classification

1) Space Configuration

The main difficulty of Space Configuration algorithms is to use appropriate parameters to represent motion or environmental space , and the degree of spatial discretization needs to be well weighed.

The main ideas of this type of algorithm:

1. Sampling or discretizing the evolution space;
2. Remove infeasible or collision areas;
3. Use the remaining parts as free-space constraints; or use pathfinding algorithm or curve planning.

a) Sampling-Based Decomposition

There are two main ways:

1. Sampling in the evolution space based on kinematic constraints requires a higher computational cost;
2. Random sampling is faster to calculate, but it depends on the sampling distribution and is unreproducible and incomplete.
   Neither is optimal and completeness cannot be guaranteed.

However, the sampling method can be used in dynamic re-planning and does not require modeling of the obstacle space.

• A more typical algorithm, PRM, introduces autonomous driving path planning in detail - a path planning algorithm based on probability sampling , mentioned by the author [33], sampling based on the reference line, and then using the objective function to find the best node set, and Speed ​​distribution on the path.
• A better strategy is to consider both the temporal and spatial dimensions ( spatiotemporal sampling ) [43].
• However, the above methods are rarely used on structured roads and are mostly used in unstructured environments.

b) Connected Cells Decomposition

Cells Decomposition first divides the space into several cells using geometric information , and then constructs the corresponding topological structure .

Figure 5 shows related applications of various Cells Decomposition. Mainly include the following two types of methods (obstacle representation method or representation not based on obstacles):

• Occupancy grid (occupancy grid) : The grid considers obstacle information and uncertainty; however, it requires a large storage space and a large amount of calculation, and is prone to errors when encountering dynamic obstacles.

• Connectivity graph : nodes represent cells and edges represent relationships between cells.

• Obstacle representation method :

  • convex polygon [45], rectangle [46], triangle [47], circle [48], [49], [50], ellipsoid shape [32], [51]; or the entire shape based on obstacle size and velocity; Road lane[52]
  • The probabilistic occupancy method was adopted in [53] .

• Not based on obstacles :

  • The most intuitive method is exact decomposition , Figure 5(e). MDP was used in [50].
  • Curvilinear or polar grids that took into account non-holonomic constraints were eliminated due to their large computational complexity and limited accuracy.
  • Vector Field Histograms (VFH) . [56];[57] considered kinodynamic constraints. VFH is more robust to uncertainty.
  • DW dynamic window method. Reduce the search space and reduce the amount of calculation, allowing dynamic obstacle avoidance.
  • Voronoi decomposition : Disadvantages: inconsistent cell sizes, lack of kinematic feasibility, not suitable for dynamic scenes, and equal spacing cannot guarantee safety. After using appropriate methods, there is still the problem of difficulty in adapting to dynamic scenes.
  • visibility decomposition：图5(a);
  • Driving corridor : Figure 5©, using road boundaries and the spatio-temporal locations of obstacles. [49] used Velocity Obstacles to construct dynamic driving corridors. Disadvantages: long calculation time.

To summarize: Cells Decomposition is still not a mainstream method in highway planning, because although it uses obstacle information, it ignores the relevant domain knowledge of the road; for highway planning, dynamic obstacle constraints and road constraints need to be combined .

c) Lattice Representation

• Lattice can usually be generated based on grids, and related motion primitives can be defined to connect connected states (states of control space or states of state space).
• Lattice is mainly used for predictive planning mentioned above
• Advantages: Consider vehicle dynamics constraints and space-time constraints; can be calculated offline for rapid re-planning
• Disadvantages: The structure is fixed and difficult to apply to reactive planning. It feels like reactive obstacle avoidance and other functions can be achieved through relatively high-frequency re-planning.

The original lattice is based on the maximum turn strategy ; later there is a maximum turn strategy that takes the speed factor into consideration . However, the predefined sets of motions provided by these two methods are often difficult to adapt to dynamic environments and require high density to achieve the desired position. Later, environment-adapted lattice appeared, which scattered points based on reference lines and lane lines; there were other methods of integrating driver behavior.

Lattice takes road boundaries and kinematic constraints into account and can achieve faster re-planning, which is very beneficial for Highway Planning. However, the storage space required for each iteration will have an impact on the amount of calculation.

2) Pathfinding Algorithms

• Pathfinding-like algorithms usually require a cost function for optimization.
• These algorithms are widely used and can be used in known or unknown environments.
• Main disadvantages : Depends on the size and complexity of the graph; requires information about space configuration, poor adaptability to open scenes.
• Introducing the classic Dijkstra, AStar and other algorithms, please refer to [Mobile Robot Motion Planning] 01 - Common Map Basics | Graph Search Basics
• 
• Examples of heuristic function design : design based on distance from lane boundary and driving distance; design based on driving time and penalties for dangerous driving.
• Anytime Weighted A * ( AWA *): weight the heuristic function to reduce calculation time.
• hybridstate A *: considering dynamic constraints, using a heuristic function based on non-holonomic constraints but not considering obstacles and a heuristic function based on obstacles but not considering non-holonomic constraints.
• D *: Based on dynamic scenes
• RRT can generate kinematically feasible paths and can be used for reactive obstacle avoidance path generation.
• RRT *: Asymptotic optimal. Disadvantages: Random sampling may lead to poor connectivity of the graph and is difficult to reproduce. For highways, the environmental information is usually relatively simple and knowable , and it may be better to use a deterministic search algorithm.

3) Attractive and Repulsive Forces

• Advantages of methods based on gravity and repulsion (fields) :

  • Can be adapted to dynamic scenes
  • No explicit modeling of the scene required

• This type of method is usually based on the gradient descent method (without considering the vehicle model) or based on Newton's second law (taking the vehicle model into account)

• Because it is difficult to model all spaces, this type of method is usually used to adjust motion trajectories in response to dynamic changes.

• 

• Artificial Potential Field (APF) : [99]; [100] Based on the idea of ​​electric fields; APF directly generates control inputs [96], or generates constraints required for optimization problems [46], [101]

• Velocity Vector Field (VVF) : Velocity vector field, more stable dynamic trajectory

• Disadvantages: Overshoot will occur when it is too close to an obstacle; it will fall into a local minimum. Random heuristics are usually used to jump out of the local minimum. However, in Highway Planning, local minima may be needed to avoid untimely lane changes.

• Virtual Force Field (VFF) : uses the vector histogram method and uses probability distribution as a virtual force.

• Another shortcoming: only the spatial dimension is considered for dynamic obstacle avoidance, so when facing obstacles, the strategy of driving by is often adopted.

• elastic band : The elastic band algorithm is improved for the above situation

Although this type of algorithm may require a large amount of calculation, it is more effective for Highway Planning.

4) Parametric and Semi-parametric Curves

• The reasons why parametric curves and semi-parametric curves are mainstream Highway Planning geometric algorithms:

  • The highway itself is a series of simple, predefined curves;
  • Collections of predefined curves are readily available as candidate test traces.
  • Kinematic constraints can be directly considered

• Geometric constraints and dynamic constraints are usually considered separately, in other words: path planning and velocity planning are decoupled [106], [107]. Decoupling enables vehicles to adapt to dynamic scenarios.

• The author mainly discusses two types of curve algorithms:

  • Directly using curves to generate the corresponding candidate trajectory library requires further selection.
    • The first type of algorithm usually uses the principle of the tentacles algorithm [108]. Compared with the space decomposition method, the calculation space and time are shortened a lot.
    • Dubins curve: combination of curve and straight line [110];
    • [42],[109]
    • However, this type of method suffers from the discontinuity of the second-order derivative of the curve curvature, which makes it difficult to apply it to vehicle models (it can be compensated by optimization or fitting methods). The author mentioned using spiral curves to connect straight lines and arc curves to achieve continuity of the second-order derivative of the curve curvature .
    • The curvature of a clothoid curve is proportional to the abscissa of the curve [50]. This feature can be used to limit lateral acceleration and constrain comfort. [111],[112],[113].
    • Clothoid curves require constant iteration, which consumes a lot of calculation time.
    • sigmoid or S-function
  • Generate curve trajectories based on path points
    • Point-based curve generation can adapt well to environmental constraints and dynamic constraints, and can also be used as a smoothing step.
    • [29]: Cubic spiral curve
    • [80]: Cubic functions are used for velocity generation, and fourth- and fifth-order polynomials are used for transverse and longitudinal trajectory generation;
    • [81]: The cubic polynomial function and the smoothed fourth-order curvature polynomial are used to generate paths, and the cubic polynomial is used to generate speed.
    • For problems with more complex polynomial interpolation calculations, semi-parametric spline curves have been improved. The curve needs to be defined as a series of piecewise polynomials.
    • Bézier curves : Use control points rather than interpolation points as inflection points , so not exactly following predefined control points. It is easy to implement and requires little calculation. [49], [116]; [117] used SVM to provide control points.
  • 
  • Summary: Curve-based algorithms are usually combined with other algorithms, and the choice of curve depends on the specific problem and environmental information.

5) Numerical Optimization

• Optimization algorithms are also part of competitive combinatorial operations research to avoid combinatorial explosion.
• The author introduces the application of two optimization algorithms in motion planning:
  • Solve complex problems while reducing search time with efficient methods.
  • Using mathematical methods to solve problems within given constraints is also the main aspect discussed in this article.
• Advantages : It is easy to handle the constraints of the problem, can handle multi-criteria optimization problems, and takes into account the state dimensions and kinematics of the vehicle model.
• Linear Programming (LP) : Linear programming. [81]
• Sequential LP (SLP) : Sequential linear programming[119]
• Nonlinear Programming (NLP) : Nonlinear programming; NLP is used in special cases of nonlinear regression problems, such as the Levenberg-Marquardt algorithm [29], [104] used in path optimization, or for the solution of the BVP problem [33] .

  • The Levenberg-Marquardt algorithm is a nonlinear least squares optimization algorithm that is often used to fit curves or solve nonlinear optimization problems. This algorithm combines the advantages of gradient descent and Gauss-Newton methods, can adaptively adjust the step size during the iteration process, and has the characteristics of fast convergence, efficiency and stability. It is widely used in curve fitting, signal processing, machine learning and other fields.

• Quadratic Programming (QP) : Quadratic programming.
• sequential quadratic programming (SQP) : sequential quadratic programming. [18]
• Mixed Integer Quadratic Programming (MIQP):[120],[121]
• Model Predictive Control (MPC) : Model predictive control [46], [47], [51], [113], [118], [122], [123]. MPC has strong re-planning capabilities, but it is deficient in solving non-convex problems and high-complexity problems.
• Dynamic Programming (DP) : Dynamic programming breaks down complex problems. [49],[124]

6) Artificial Intelligence

• AI algorithms have strong generalization capabilities and can make new adjustments without changing the algorithm structure.
• Four types of algorithms: humanoid, heuristic, approximate reasoning, logic

a) AI Logic-Based Approach

• Expert systems are defined to solve specific complex tasks and rely on knowledge bases and reasoning engines to automate the reasoning system.

• Logic-based AI algorithms have the advantage of fast architecture (not well understood) and can be used for predictive or reactive planning.

• Main advantages : Can simulate human logic and rational reasoning

• Disadvantages : The knowledge base requires discretization of environmental variables with a large number of causal rules and tuning parameters.

• rule-based：

  • [126] achieved satisfactory results for rule-based intent prediction on highways and used it for lane changing maneuvers in [127].
  • Its main advantages are that causal relationships can be clearly identified, labeling is convenient, and implementation is simple.
  • Its main disadvantage is circular reasoning and exhaustive enumeration of rules, leading to infinite loops and affecting calculation time; if the current situation does not match the observations in the rule base, inappropriate decisions may be made.
  • The order of the rules is also important, and rules can be ordered heuristically .

• decision treesdecision tree :

  • [78] Describe all possible lanes of travel;
  • [32], [128] use binary decision diagrams and flow charts to represent Boolean functions.
  • Difficulty describing uncertain and approximate situations. The decision tree should also consider the dangerous and illegal behavior of other cars, that is, consider legal factors.

• Finite State Machine (FSM) finite state machine:

  • [55] is used for emergency parking; [129] maps the sequence of actions and states into path generation and tracking control; [18] uses parallel state charts to describe simultaneous states.
  • The main disadvantage is that they are only based on certain knowledge and cannot be generalized to unknown situations.
  • 

• Bayesian networks Bayesian networks:

  • Allow for uncertainty.
  • Bayesian networks are statistical representations of causal relationships based on probabilistic transformations. They proposed a knowledge-based parameter identification step to determine the most likely sequence of states in the output sequence.
  • [27] Compare the intentions of obstacles and return the operation of the own vehicle.
  • [50] adopted MDP to select trajectories in the tentacle trajectory library. [130] used MDP in lane keeping and lane changing decisions.
  • The advantage of MDP is the ability to evaluate multiple prediction information simultaneously.
  • POMDP:[131],[132],[133]

• Summary :

  • Methods based on artificial intelligence logic are mostly suitable for constrained and predictable environments (such as high-speed scenarios)
  • The main disadvantage is the lack of autonomy and strict program structure, and the inability to implement adaptive and reconfigurable algorithms.

b) AI Heuristic Algorithms

• Heuristic algorithms are based on experience and facilitate the exploration of natural environmental processes. Their goal is to find an approximate solution, so they can be used as faster and more efficient algorithms when traditional exhaustive methods fail.
• Usually returns a series of action collections.
• Their main advantages are short calculation time, low complexity, and strong ability to handle complex problems.
• However, global optimality and accuracy cannot be guaranteed .
• Agent : Heuristic algorithms usually involve the concept of an agent.
  • The functionality of the heuristic agent allows the inclusion of multi-strategy decisions, such as [132], including distance to target, lane selection bias, maximum yaw rate and policy cost.
  • Adapt to uncertain environments
• Support Vector Machines (SVM)：
  • [117] used SVM to fit the control points of the Bezier curve; [134] was applied to lane changing decisions.
• Evolutionary methods ：
  • Evolutionary methods Evolutionary algorithms are a type of widely used learning algorithms inspired by biological simulations and have natural and reasonable evolutionary processes, such as reproduction, mutation, recombination and selection . First, a set of a priori solutions associated with the fitness function needs to be identified to evaluate their quality. A set of evolutionary processes is then applied to find better solutions to the optimization problem.
  • [136] applied two genetic algorithms to improve the fuzzy control module, and [137] introduced in detail a 4-chromosome structure genetic algorithm based on speed, angle, interruption and avoidance time.
  • The stability and risk of this algorithm are related to the mutation mechanism , whose random process may lead to local minima .

c) AI Approximate Reasoning

• The difference from logical reasoning is that the knowledge base is non Boolean; and the learning method of classifying new knowledge to adapt to future situations .

• fuzzy logic：

  • Fuzzy logic relies on multi-valued variables and includes fuzzy expert systems based on Boolean trade-offs that can mathematically model fuzziness and approximate cognitive reasoning.
  • The main advantage is the flexibility and tolerance of the designed rules and their scalability to uncertain data
  • The main disadvantages are the lack of traceability and the lack of a systematic design approach.
  • [138] Determine the lane change timing based on the distance to the obstacle. [139] and [140] proposed automatic overtaking maneuvers and lateral following based on reference maps using fuzzy rules for lateral errors and angular errors respectively.
  • 

• Artificial Neural Networks (ANN)：

  • ANN has applications in decision-making and path generation.
  • The main advantage is their ability to learn by training on multidimensional data
  • Lack of a good causal explanation of the solution.
  • Another disadvantage is the large amount of data required , which makes them computationally expensive .
  • [122] adopted a method combining Hidden Markov Models ( HMM ) hidden Markov models and Gaussian mixture regression.
  • [141] used the Levenberg-Marquardt algorithm on a three-layer artificial neural network to learn driving speed.
  • [142] used a convolutional neural network to simulate the driver's lane change trajectory, while [143] compared the learning phase with the control input of a human driver.
  • Reinforcement learning is used in [133] to determine the strategy of each agent in the scene representation, and [144] and [145] are used in multi-target overtaking and automatic lane-changing maneuvers.
  • Deep learning methods were explored in [146], while [147] integrated sensing and planning blocks into end-to-end learning.
  • In [148], Bayesian probability is used to extend the neural network and perform belief networks learning under uncertainty .

• Summary :

  • AI algorithms for approximate reasoning rely on logical and statistical foundations that extend to cognitive properties. They also provide adaptive reasoning to evolve appropriately for their environment. Good adaptability and generalization ability to complex changing environments.
  • The main drawback of these algorithms is the lack of feedback from real driving scenarios.

d) AI Human-Like Methods

• The main advantage is the abstract and universal representation of the decision-making process

• But difficult to model and use accurately

• Three types of methods: risk, task, and game

• Risk estimators：

  • Risk estimators are used for sound decision-making in scenario analysis with cognitive biases.
  • Advantages: Intuitive.
  • Related evaluation indicators:
    • Time-To-Collision (TTC):[51],[55],[150];
    • Time-To-React (TTR):[151],[152]
    • Time-To-Intercept (TTI)：[154]
  • Binary risk estimators Binary risk estimators provide a rough assessment of the scenario, but the uncertainty of the future motion of the obstacle is a parameter that cannot be ignored.
  • Probabilistic risk estimators Probabilistic risk estimators are better suited to providing a more realistic representation of scenarios.
  • compensation risk：
    • Compensation risk: To adapt to the driver's risk perception, the compensation risk estimator introduces the driver's state, such as stress, sleepiness, or illness. Time of Zone Clearance (TZC) was introduced in [157].

• Task， taxonomic models：

  • They identify sequence-driven task relationships through behavioral and competency requirements.

  • consists of a series of tasks

  • One disadvantage is that they are difficult to describe explicitly because they either focus on driving tasks such as changing lanes or merging, or they are difficult to be exhaustive.

• game theory game theory:

  • The main advantage of this approach is that it quickly obtains a trained driver model
  • The disadvantage is that it assumes that all participants follow the rules , but does not take into account some of the variations in real-world situations.
  • In [133], the model develops more complex policies when trained on other behaviors. The authors use POMDP to model the knowledge of traffic participants and Jaakola reinforcement learning during the training phase .

IV. COMPARISON TABLE FOR HIGHWAY APPLICATIONS

The table summarizes previous algorithms.

TABLE II COMPARISON TABLE FOR HIGHWAY APPLICATIONS OF MOTION PLANNING METHODS (‘−−’ VERY INAPPROPRIATE, ‘−’ INAPPROPRIATE,‘∼’ INTERMEDIARY, ‘+’ APPROPRIATE, ‘++’ VERY APPROPRIATE).

V. FUTURE RESEARCH DIRECTIONS

A. Data Management

• Reference [166] details several perceptual challenges affecting motion planning, and Reference [167] discusses conditions for improving driving control algorithms.
• The author focuses on two issues in this part:
  • How much data is enough in terms of uncertainty and credibility? how much data is enough in terms of uncertainties and trustworthiness?
  • How do we make decisions about delayed and asynchronous data? how do we make a decision with latency and unsynchronized data?

1) Data Uncertainty and Trustworthiness

• Algorithms related to data uncertainty and credibility:
  • Kalman estimatorsKalman estimators [52], [169]
  • Markov processes Markov processes [131], [170]
  • Monte Carlo method Markov processes [131], [170]
  • evidential theory [52]
  • interval algorithm interval arithmetic [171]
• In addition, the planning phase must ensure movement continuity to ensure vehicle safety even if data is lost.

2) Data Synchronization

• The data for the scene is obtained with different frequencies and delays
• The first solution is to run the algorithm with the same time step for the motion planning module and the controller and extrapolate the last obtained data.
• The second approach is based on a planner that is refreshed every time new data arrives. The hybrid logic will then run at a fixed frequency and refresh the data as new data arrives, improving the propagation equation by extrapolating the data set.

B. Adaptive Mobility

• It points out the safety issues that human drivers need to consider when facing robot drivers, including respect, acceptability, predictability or comfort.

1) Safety

• For security reasons, algorithms smart enough to distinguish between permanent or temporary dangers are needed
• Longitudinal distance and speed control in autonomous driving are usually faster than human driving, so the longitudinal safety space can be smaller
• In comparison, the lateral distance control of human drivers is usually more stable; therefore, the lateral safety space of autonomous driving must be larger.
• In addition, this safe space must be sufficient to cope with accidents in the environment (e.g., unexpected braking of the vehicle ahead, unforeseen dead ends).
• Issues related to human takeover and human-machine hybrid driving of self-driving cars :
  • The first issue is maintaining stable and predictable decision-making in situations where the driver takes back control
  • The second is that the decisions of the human driver and the machine may conflict with each other [177].
  • Safety issues that human drivers need to consider when facing autonomous driving include aspects such as respect, acceptability, predictability or comfort.

2) Eco-driving Planning

• Eco-driving is expressed as a multi-criteria optimization problem (see III-B5), with a cost function based on energy efficiency and environmental protection, to achieve the goal of energy saving and low carbon. There are two main aspects to consider.
• First, the dynamics model of autonomous vehicles was analyzed, and methods for smooth maneuvers such as low acceleration and constant speed were proposed to reduce fuel consumption.
• Second, have good predictions and expectations to avoid sudden maneuvers

3) Perception Compensation

• One issue regarding perception is the constraint of a fixed perspective of the driving scene
• However, when a human driver has to make a decision, he/she optimizes his/her perception by moving the vehicle to gain more information in the larger perceived environment.

4) Route Context

• Global path planning actually creates corresponding constraints for local path planning. For example, a right-hand highway exit requires self-vehicle navigation to the rightmost lane.

5) Service Orientation

• According to the use of different vehicles, the decision-making behavior is adjusted based on service attributes.
• The classic approach so far is spatiotemporal domain decision-making, i.e. when and where the vehicle must navigate. Now adds a third dimension: how the vehicle navigates

C. Cooperative Planning

• individual：

  • Each vehicle makes decisions independently of other vehicles but takes into account factors such as interaction [159] and politeness [184].

• collective：

  • a swarm intelligence : swarm intelligence
  • This collective decision-making requires knowledge of the status of other vehicles and strong communication. Through swarm intelligence, event-driven methods, risk estimation and fuzzy logic modeling, vehicles can be enabled to collaborate and achieve formation maneuvers. At the same time, smart infrastructure can also serve as a manager or controller to realize the anticipation and monitoring of vehicle actions through a coordinated scheduler. The literature mainly introduces smart highway managers based on knowledge reasoning engines and agents.

• shared：

  • How autonomous driving technology works with human drivers and makes shared decisions. SAE automation levels describe the shared cooperation between the driver and the autonomous vehicle. The driver's experience is useful for the autonomous driving system, and fuzzification methods can be used to combine it with system decisions. The research also explores how to adapt the behavior of autonomous vehicles to the type of driver to better promote autonomous driving technology.

D. Validation and Evaluation

1) Transition Stability

• “Transition Stability” refers to transition stability, that is, whether transitions between systems are stable, reliable, and robust when using different motion planning algorithms. If the transition is unstable, a high-level supervisor needs to be added to verify the consistency of decisions, actions, and observations.
• In [195], the authors proposed a behavioral model similar to FSM decision-making, implementing an architecture that can switch between manual and automated driving and between different maneuvering modes.

2) Evaluation

• It mainly includes human driver evaluation (safety, smoothness, operating time) and the use of key performance indicators (KPIs) in simulation and experimental tests.

3) Ethics

• Ethical and moral issues need to be considered when robotic systems interact with humans. Related research proposes strategies for deploying ethics based on hybrid rationality and artificial intelligence, as well as methods for incorporating ethics into decision-making.

4) Algorithm Relaxation

• Strict compliance and laxity of driving rules are key to ensuring the safety of self-driving cars. In fact, even hard constraints can become dangerous in critical situations [13].