PaperlayerView exact revision on arXiv

Source-linked AI summary

A Systematic Survey of Control Techniques and Applications in Connected and Automated Vehicles

Wei Liu, Min Hua, Zhiyun Deng, Zonglin Meng, Yanjun Huang, Chuan Hu, Shunhui Song, Letian Gao, Changsheng Liu, Bin Shuai, Amir Khajepour, Lu Xiong, Xin Xia

arXiv:2303.05665v2cs.RO

TL;DR

Vehicle control depends on accurate state information, yet important states such as sideslip angle cannot be directly measured on commercial vehicles. This survey reviews sideslip estimation, AV trajectory tracking, and CAV collaborative control, concluding that these control methods have potential to improve transportation efficiency, safety, and sustainability.

  • Problem

    Important vehicle states such as sideslip angle cannot be directly measured in commercial vehicles, while existing reviews have not cohesively connected state estimation with vehicle control for AVs and CAVs.

    Source: §I · pp. 1–2

  • Method

    The survey reviews sideslip-angle estimation, AV trajectory-tracking control, and CAV collaborative-control techniques, applications, and enabling methods.

    Source: §I · p. 3, §V · p. 19

  • Results

    The survey synthesizes estimation approaches, three types of AV trajectory-control algorithms, and CAV collaborative-control methods and applications.

    Source: §V · pp. 18–19

  • Takeaways & Limitations

    The reviewed cooperative driving-control methods have the potential to improve transportation efficiency, safety, and sustainability.

    Source: §V · p. 19

Abstract

from arXiv · show

Vehicle control is one of the most critical challenges in autonomous vehicles (AVs) and connected and automated vehicles (CAVs), and it is paramount in vehicle safety, passenger comfort, transportation efficiency, and energy saving. This survey attempts to provide a comprehensive and thorough overview of the current state of vehicle control technology, focusing on the evolution from vehicle state estimation and trajectory tracking control in AVs at the microscopic level to collaborative control in CAVs at the macroscopic level. First, this review starts with vehicle key state estimation, specifically vehicle sideslip angle, which is the most pivotal state for vehicle trajectory control, to discuss representative approaches. Then, we present symbolic vehicle trajectory tracking control approaches for AVs. On top of that, we further review the collaborative control frameworks for CAVs and corresponding applications. Finally, this survey concludes with a discussion of future research directions and the challenges. This survey aims to provide a contextualized and in-depth look at state of the art in vehicle control for AVs and CAVs, identifying critical areas of focus and pointing out the potential areas for further exploration.

I. INTRODUCTION

The survey links vehicle state estimation and trajectory control for AVs with collaborative control for CAVs. It focuses on sideslip-angle estimation because this difficult-to-measure state is important to vehicle control, then reviews the surrounding methods and applications.

I. INTRODUCTION

  • Accurate state information is a prerequisite for vehicle control, but key states including sideslip angle cannot be directly measured in commercial vehicles.Sideslip angle combines longitudinal and lateral velocity information, and errors in other states such as longitudinal velocity and roll can affect its estimate.Source: §I · p. 1

II. VEHICLE SIDESLIP ANGLE ESTIMATION

  • Sideslip angle is difficult to estimate because errors in other states contribute to its error, yet it is critical to stabilization, motion planning, road-condition estimation, handover, and navigation.The survey focuses on sideslip angle estimation using information from onboard sensors such as SWAS, WSS, IMU, and GNSS.Source: §I · p. 1

I. INTRODUCTION

  • The survey addresses a review gap by treating state estimation and AV and CAV vehicle control cohesively, rather than discussing AVs and CAVs separately.It also notes that existing reviews tend to focus on CAV control in specific scenarios and may overlook the importance of state estimation.Source: §I · p. 3
  • The survey frames state estimation, AV trajectory tracking, and CAV collaborative control as successive parts of its review, with future research directions addressed at the end.It covers physical vehicle models and tracking algorithms, then CAV enabling techniques, methodologies, and applications.Source: §I · p. 3

II. VEHICLE SIDESLIP ANGLE ESTIMATION

  • Sideslip-angle estimation is organized into onboard-sensor-based, GNSS-augmented, and artificial-intelligence-augmented approaches.The onboard-sensor-based group includes vehicle-kinematics, vehicle-dynamics, and dynamic-kinematics-fusion approaches; GNSS-augmented methods include state-based and state-error-based approaches.Source: §I · p. 2, §II · p. 4

A. OSB approach

Onboard-sensor-based sideslip estimation uses vehicle kinematics, vehicle dynamics, or their fusion. Vehicle-dynamics-based methods offer a model-driven route, but their accuracy is closely tied to the reliability of the vehicle model and its parameters.

A. OSB approach

  • Onboard-sensor-based estimation is categorized into vehicle-dynamics-based, vehicle-kinematics-based, and dynamic-kinematics-fusion-based approaches.The survey introduces the vehicle-dynamics-based approach through a classical vehicle model and its variants.Source: A. OSB approach · p. 4, 1) VDB approach · p. 4

1) VDB approach:

  • Vehicle-dynamics-based estimation performance depends strongly on vehicle-model reliability, which can be affected by degrees of freedom, road friction, tire nonlinearity, and parameter uncertainty.Parameters such as center of gravity and tire cornering stiffness vary with driving and load conditions; uncertainty can cause estimation error or prevent convergence.Source: 1) VDB approach · pp. 4–5
  • Extended Kalman filters are widely used with vehicle-dynamics models, with variants addressing noise assumptions, tire models, sensor measurements, and roll effects.Examples include adaptive noise updates, interacting multiple-model filtering, added lateral tire-force measurements, and a 4-DOF model incorporating roll motion.Source: 1) VDB approach · p. 5
  • When tire behavior enters an extreme nonlinear region, UKF avoids EKF’s local linearization by using sigma points, and cited studies report better sideslip estimation than EKF.Other approaches include constrained UKF, particle filters, cubature Kalman filters, and nonlinear observers.Source: 1) VDB approach · p. 5

2) VKB approach:

Vehicle-kinematics-based estimation uses sensor information without relying on an accurate vehicle-dynamics model, while dynamic-kinematics fusion combines the two approaches. The survey describes integration drift as a key kinematic challenge and condition-dependent fusion as a way to use the approaches complementarily.

2) VKB approach:

  • Vehicle-kinematics-based estimation uses sensors such as the IMU, wheel-speed sensors, and steering-angle sensors to avoid dependence on potentially inaccurate vehicle-dynamics models.Direct integration can accumulate error or diverge; an EKF and additional processing of speed, sensor offsets, roll, and center-of-gravity acceleration are among approaches described.Source: 2) VKB approach · p. 5
  • IMU-centered kinematic estimates require correction of temperature drift, bias, noise, and gravity components, since prolonged acceleration integration can accumulate error.Gravity effects arise from roll and pitch angles.Source: 2) VKB approach · p. 5

3) DKFB approach:

  • Dynamic-kinematics fusion combines VDB estimates, which are more confident at low frequencies, with VKB estimates, which perform better at high frequencies.The survey describes a first-order filter for combining the approaches and reviews further fusion strategies using weighting and driving-condition information.Source: 3) DKFB approach · pp. 5–6
  • A reviewed DKFB architecture uses VDB feedback under normal driving conditions but switches to VKB alone in critical conditions.Its performance is verified in double-lane-change and slalom maneuvers.Source: 3) DKFB approach · p. 6

B. GAU approach

GNSS-augmented sideslip-angle estimation methods use GNSS position and velocity information and divide into state-based and state-error-based approaches. The survey reviews their designs, limitations, and links to vehicle navigation and CAV collaborative control.

B. GAU approach

  • GNSS-augmented (GAU) sideslip-angle estimation incorporates GNSS position and velocity information and comprises state-based (SB) and state-error-based (SEB) approaches.The review focuses on these approaches because GNSS is commonly used in AVs and CAVs.Source: B. GAU approach · p. 6

1) SB approach:

  • SB methods directly use model state variables and are simple and computationally inexpensive, but GNSS sampling and sensor synchronization can constrain estimation.Reviewed work uses EKFs or nonlinear observers, multisensor fusion, and methods addressing low-rate GNSS, delayed measurements, and integrated-error accumulation during GNSS outages.Source: 1) SB approach · p. 6

2) SEB approach:

  • SEB methods fuse GNSS and IMU measurements through an INS error-state framework, using position and velocity errors and their covariances as KF measurements.Their state vectors typically include position, velocity, and attitude errors plus gyro and acceleration biases; the review reports improved attitude-estimation accuracy and robustness over SB methods.Source: 2) SEB approach · p. 7, Figure 4 · p. 7
  • Sideslip-angle estimation remains challenging because of observability, while sideslip-informed vehicle navigation can improve position information relevant to CAV collaborative control.Reviewed SEB work addresses heading error and GNSS sampling limits through methods including VDB augmentation, dual-antenna GNSS, and fusion weighted by lateral excitation.Source: 2) SEB approach · p. 7

C. AIA approach

The survey reviews data-driven and hybrid artificial-intelligence approaches to sideslip-angle estimation, then places vehicle trajectory tracking in the AV control pipeline. It describes tracking methods and vehicle-model choices, including their accuracy and complexity trade-offs.

C. AIA approach

  • Pure artificial-intelligence-augmented (AIA) methods estimate sideslip angle end to end, while hybrid AIA combines mathematical vehicle models with AI techniques.Pure approaches include neural networks trained for different road conditions; hybrid examples feed LSTM outputs into dynamic EKF/UKF models or combine neural networks with a kinematic-model UKF.Source: 2) Hybrid AIA approach · pp. 7–8

2) Hybrid AIA approach:

  • Hybrid AIA can improve estimation-method interpretability to a certain extent compared with pure AIA.Examples combine LSTM with dynamic EKF/UKF models or an artificial neural network with a kinematic-model UKF.Source: 2) Hybrid AIA approach · p. 8

III. TRAJECTORY TRACKING CONTROL OF AVS

  • AV trajectory tracking control computes steering, throttle, and braking commands to follow a predefined path while accounting for vehicle dynamics and environmental uncertainties.The survey groups methods into feedback control without prediction, predictive feedback such as MPC, and learning-based control such as deep reinforcement learning.Source: §III · p. 8

A. Vehicle mathematical model

  • Vehicle tracking uses VKB, bicycle VDB, or four-wheel VDB models; greater model accuracy can increase complexity and computation.Pure pursuit uses a look-ahead distance to calculate lateral error, whereas the Stanley method considers both lateral and heading errors and requires more parameter tuning; VKB ignores tire slip, which matters in critical driving conditions.Source: A. Vehicle mathematical model · p. 8

B. Feedback control algorithms without prediction

The section surveys non-predictive feedback control for AV trajectory tracking, focusing on PID and LQR methods, their applications, and practical constraints.

B. Feedback control algorithms without prediction

  • Non-predictive feedback methods such as PID, LQR, and SMC are reported to provide robust trajectory tracking under most driving conditions.The section classifies these methods as feedback control without prediction.Source: B. Feedback control algorithms without… · p. 8

1) PID control:

  • PID variants adapt gains to road curvature, filter steering for smoothness, or incorporate yaw rate, but gain tuning remains difficult as driving conditions change.Examples also include nested PID steering and parameters optimized by PSO.Source: 1) PID control · p. 9

2) LQR control:

  • LQR minimizes a quadratic cost and obtains feedback gain through a Riccati equation; vehicle tracking commonly uses dynamics models, with Q and R tuned for convergence and stability.Researchers also combine LQR with disturbance compensation or LMI-based designs for tracking and critical driving conditions.Source: 2) LQR control · pp. 9–10
  • LQR assumes small control inputs, so sharp curvature or low adhesion can compromise robustness, while linearization of rapidly varying nonlinear dynamics can distort the model.The survey notes these limitations motivate exploring more advanced vehicle-control techniques.Source: 2) LQR control · p. 10

3) SMC control:

The section presents sliding mode control as a robust nonlinear approach to AV trajectory tracking, while emphasizing chattering and real-time implementation trade-offs.

3) SMC control:

  • SMC addresses nonlinear tracking with parametric uncertainty and external disturbances, but its super-twisting plane must ensure stability while reducing chattering.The section describes the sliding-mode objective as convergence to the surface s = 0.Source: 3) SMC control · p. 10
  • The SMC formulation models sliding-variable dynamics using state and control terms, then generates a steering-angle signal under stated bounds and parameter conditions.The prose identifies f and g as continuous functions, and x and u as the state vector and control input.Source: 3) SMC control · p. 10
  • AV tracking studies adapt SMC gains or sliding surfaces and combine SMC with optimization or fuzzy logic to address parameter variation, tracking errors, and jittering [166–171].High-order sliding modes reduce chattering but add computation; second-order sliding mode is described as the mainstream choice for real-time performance.Source: 3) SMC control · pp. 10–11
  • SMC is less dependent on model parameters, but chattering remains problematic in real-time use and nonsmooth references can produce high-frequency control-input oscillations.The survey also notes hybrid schemes combining SMC with PID or LQR.Source: 3) SMC control · p. 11

4) Other control methods:

The section reviews other AV trajectory-tracking methods and MPC, describing their control strategies, applications, and implementation challenges.

4) Other control methods:

  • H-infinity, adaptive, and fuzzy control address robustness, nonlinearities, steering backlash, transient response, or representation of human expertise in AV tracking.Examples include robust H-infinity strategies for environmental uncertainty and adaptive compensation for steering backlash.Source: 4) Other control methods · p. 11
  • Highly nonlinear, cross-coupled vehicle dynamics and unpredictable traffic environments remain challenges for robust AV trajectory tracking.Source: 4) Other control methods · p. 11

C. Feedback control algorithms with prediction

  • MPC predicts future behavior over a finite horizon, optimizes a constrained control sequence, applies only its first input, and repeats the process iteratively.The approach can account for disturbances through constraints, at higher computational effort than classical tracking methods.Source: C. Feedback control algorithms with… · p. 11, Figure 6 · p. 11
  • MPC studies target optimization feasibility, closed-loop stability, and robustness, while reported designs combine models and controllers for obstacle avoidance, stability, and difficult maneuvers.The survey notes that model accuracy, horizon limits, computational complexity, system delays, and optimization initialization can impede real-time use.Source: C. Feedback control algorithms with… · p. 12

D. Learning-based control algorithms

The survey reviews learning-based approaches to trajectory tracking, including methods that supplement or replace model-based control. It also notes their practical limitations and the dependence of algorithm choice on application requirements.

D. Learning-based control algorithms

  • Sampling-based MPC, self-learning MPC, and competitive control are presented as learning-based ways to optimize MPC [200, 201, 202, 203].The survey characterizes self-learning MPC as closer to the optimal control rate based on known parameters.Source: D. Learning-based control algorithms · p. 12
  • Deep reinforcement learning may address trajectory-tracking limits of modern control algorithms while avoiding dependence on a specific vehicle model.The survey notes that vehicle models are difficult to develop because of strong nonlinearity and uncertainty.Source: D. Learning-based control algorithms · p. 12
  • Examples include actor–critic networks, PPO-trained agents, PAC algorithms using online data, and MEDIRL architectures that scale with environment and dataset complexity.The methods respectively approximate critic and actor networks, train in simulation, reduce dependence on specific approximation structures, and extend scalability.Source: D. Learning-based control algorithms · p. 12
  • Zhang et al. optimize a reinforcement-learning residual policy around a modified artificial-potential-field guiding controller, with experiments outperforming leading algorithms [220].Chen et al. combine PPO with a PP model, which supplies baseline steering while PPO provides corrections.Source: D. Learning-based control algorithms · pp. 12–13
  • Deep reinforcement learning can require extensive training data, overfit, and be difficult to interpret [no marker supplied].The survey describes the field as rapidly expanding despite these challenges.Source: D. Learning-based control algorithms · p. 13
  • MFAC transforms trajectory tracking into stabilization using preview-deviation-yaw angle, while another system combines extremum seeking with learning-based control.End-to-end neural networks trained on limited data tracked trajectories effectively on two TORCS tracks.Source: D. Learning-based control algorithms · p. 13
  • The appropriate trajectory-tracking algorithm depends on the application and its performance requirements.The survey presents its review as covering AV trajectory-tracking algorithms at the microscopic level.Source: D. Learning-based control algorithms · p. 13

IV. COLLABORATIVE CONTROL OF CAVS

The survey presents collaborative CAV control as a response to limits in onboard perception, using shared information for coordinated decisions. It reviews enabling techniques and system components, including vehicle models, information exchange, and communication quality.

IV. COLLABORATIVE CONTROL OF CAVS

  • Onboard sensors cannot ensure reliable perception in all dynamic conditions, while shared information beyond line of sight can support coordinated CAV decisions and improve traffic efficiency, comfort, and economy.The cited conditions include adverse weather, sensor or vehicle-model uncertainty, and illumination variation.Source: §IV · p. 13

A. Enabling techniques

  • Collaborative control uses V2V and V2I to coordinate vehicles through shared speed, position, and destination data, with some systems also sharing visual information [245] [246].Reported enabling benefits include smoother traffic flow and a greater than 20% increase in object-detection accuracy using shared visual information.Source: A. Enabling techniques · p. 13

1) Description of collaborative control systems:

  • Collaborative systems use longitudinal dynamics models for platooning, merging, and lane changes; higher-order models can be more accurate but require more computation and data.Model choice depends on required accuracy, computational resources, and data availability.Source: 1) Description of collaborative… · p. 14
  • Information-flow topologies specify how vehicles exchange information, with PF, PLF, BD, BDL, TPF, and TPLF among the listed schemes.The choice depends on movement context and goals such as reducing delays, fuel use, or accidents.Source: 1) Description of collaborative… · p. 14
  • Communication delays and packet loss can impair coordination through missing or incomplete data, with congestion, interference, and hardware failures listed as causes [255–263].Mitigations include message prioritization, routing and congestion-control methods, retransmission, error-correction codes, and network redundancy.Source: 1) Description of collaborative… · pp. 14–15

2) Objectives of collaborative control systems:

The survey distinguishes stability objectives for individual CAVs and whole platoons, then compares centralized, decentralized, and mixed coordination modes. The coordination choice depends on application requirements and available resources.

2) Objectives of collaborative control systems:

  • Individual stability means each CAV maintains desired speed and distance from neighbors, with sensing and control such as ACC or CACC supporting that objective.The controller should prevent overshooting or undershooting while accounting for noise and disturbances.Source: 2) Objectives of collaborative control… · p. 15
  • String stability concerns the whole platoon maintaining safe spacing and compensating for disturbances, using spacing policies and potentially consensus, MPC, feedback, or feedforward control.The survey notes that string stability is more challenging because it coordinates vehicles with different dynamics.Source: 2) Objectives of collaborative control… · p. 15

1) Coordination scheme:

  • Centralized control offers coordination but needs substantial communication and computation, whereas decentralized control is more scalable and resilient but may sacrifice optimal coordination.A mixed mode divides coordination between central commands and local adaptation [p43–p45].Source: 1) Coordination scheme · p. 15
  • The survey associates centralized schemes with well-defined settings, decentralized schemes with platooning, and mixed schemes with urban traffic management [p47].Selection also depends on application requirements, resources, and autonomy level.Source: 1) Coordination scheme · pp. 15–16
  • Collaborative methods must manage inter-vehicle interactions and changing conditions, while robust control of many variables remains a design challenge.They rely on communication about vehicle states and objectives.Source: 1) Coordination scheme · p. 16

2) Methodologies:

The survey groups collaborative-control methodologies into game-theoretic, optimization-based, and reinforcement-learning approaches, and describes applications organized by collaboration level. It also notes limitations involving computation, uncertainty, training data, and interpretability.

2) Methodologies:

  • Game-theoretic methods model CAV interactions as games and seek equilibria, with examples targeting energy use, collisions, travel time, and system performance.The survey notes that these methods can be computationally expensive and challenging to implement in practice.Source: 2) Methodologies · p. 16
  • Optimization-based methods seek group-wide performance objectives, including trajectory optimization for delay and joint signal-timing and trajectory control.The survey cautions that these methods may not handle real-world uncertainty and unpredictability well.Source: 2) Methodologies · p. 16
  • Reinforcement-learning methods train control algorithms through trial and error and include speed-control, mixed-autonomy, and blockchain-integrated multi-agent examples.They may need substantial training data and can be difficult to interpret and validate.Source: 2) Methodologies · p. 16

D. Applications

  • The survey organizes applications as one-, two-, and three-dimensional collaboration levels, each with distinct technical and operational challenges and opportunities [107].It notes that much existing research addresses specific scenarios or facilities such as platooning, lane changes, merging, and intersections.Source: D. Applications · pp. 16–17

1) One-dimensional collaboration:

The survey groups CAV collaborative control applications by the dimensions of vehicle motion they coordinate: longitudinal, lateral, or both with timing. These applications include platooning, lane changes, merging, and intersection management.

1) One-dimensional collaboration:

  • Truck platooning can reduce highway fuel consumption by about 10%, depending on inter-vehicle distance and speed, while synchronized braking and acceleration can reduce accident risk.Platooning also aims to reduce emissions, congestion, and improve traffic flow.Source: 1) One-dimensional collaboration · p. 17
  • For platoon formation, consensus protocols converge vehicle motion states to a common desired value, while an evolutionary algorithm optimizes longitudinal trajectories with an energy-aware objective.The evolutionary approach avoids assuming the solution landscape and requires relatively shorter runtime than programming models.Source: 1) One-dimensional collaboration · p. 17

2) Two-dimensional collaboration:

  • Collaborative lane-changing strategies use communication, game-theoretic approaches, or optimal control to coordinate maneuvers, with goals including safer changes, improved traffic flow, and individual and social benefits.Optimal-control examples jointly consider maneuver time, energy use, and safety constraints; other methods address lane drops and conflict areas.Source: 2) Two-dimensional collaboration · p. 17

3) Three-dimensional collaboration:

  • At highway on-ramps, collaborative merging coordinates vehicles to avoid collisions and reduce abrupt braking; one mixed-traffic approach reports a 10-15% increase in traffic throughput.Approaches include rule-based coordination, digital-twin advisory systems, hierarchical control, and optimization.Source: 3) Three-dimensional collaboration · pp. 17–18
  • At signalized intersections, collaborative systems coordinate vehicles with infrastructure; surveyed approaches target energy use, emissions, travel delay, throughput, fuel consumption, and intersection capacity.Examples include corridor eco-driving and joint optimization of traffic signals and vehicle platooning.Source: 3) Three-dimensional collaboration · p. 18
  • At unsignalized intersections, surveyed methods coordinate vehicle motion and entry timing using conflict graphs, bi-level optimization, or digital-twin control frameworks.The approaches aim to resolve conflicts, improve capacity and space utilization, and prevent spillbacks.Source: 3) Three-dimensional collaboration · pp. 17–18

V. CONCLUSION AND FUTURE WORKS

The survey reviews vehicle state estimation, AV trajectory tracking, and CAV collaborative control, then identifies research directions for improving robustness, mixed-traffic coordination, resilience, and safety evaluation.

V. CONCLUSION AND FUTURE WORKS

  • The survey covers sideslip-angle estimation across sensor configurations and model features, AV trajectory tracking methods, and CAV collaborative-control techniques and applications.Trajectory tracking methods are analyzed by model complexity, computation cost, optimal performance, and application scenarios.Source: §V · pp. 18–19
  • Robust state estimation requires real-time sensor-signal confidence assessment and fusion methods that account for environmental sensitivity and differing sensor latency and frequency.GNSS, LiDAR, and cameras can support estimation, but are susceptible to buildings, illumination, and weather.Source: §V · p. 19
  • Trajectory-control design should account for uncertainty in estimated vehicle states, including when extending models from a two-dimensional plane to three-dimensional space.The survey notes that multimodal information can support real-time three-dimensional vehicle-attitude estimation.Source: §V · p. 19
  • Mixed-autonomy platoons pose a challenge because human-driven behavior introduces heterogeneous vehicle dynamics that control techniques must accommodate while maintaining safety and efficiency.Suggested directions include predicting human-driver behavior and combining centralized and decentralized control.Source: §V · p. 19
  • CAV collaborative control needs resilience to sensing and communication failures and cyber-attacks, with security and privacy considered in communication, sensing, and control protocols.The survey identifies encryption, authentication, and access control as examples of relevant mechanisms.Source: §V · p. 19
  • Safety evaluation would benefit from a unified simulation platform and intelligent testing environments that model CAV interactions with vehicles and infrastructure.The proposed platform would support evaluation before real-world implementation and use machine learning for accelerated testing and training.Source: §V · p. 19

Related papers

Loading 2303.05665v2…