PaperlayerView exact revision on arXiv

Source-linked AI summary

Milestones in Autonomous Driving and Intelligent Vehicles Part I: Control, Computing System Design, Communication, HD Map, Testing, and Human Behaviors

Long Chen, Yuchen Li, Chao Huang, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Siyu Teng, Chen Lv, Jinjun Wang, Dongpu Cao, Nanning Zheng, Fei-Yue Wang

arXiv:2305.11239v2cs.AIcs.ROcs.SYeess.SY

TL;DR

Prior AD and IV surveys often focus on specific tasks and lack systematic summaries and future research directions. This article is the technical survey’s first part, reviewing milestones across six areas; together with the other parts, the work aims to guide newcomers and connect past and future research.

  • Problem

    Many AD and IV surveys focus on specific tasks and lack systematic summaries and future research directions.

    Source: §I · p. 1

  • Method

    This article reviews milestones and development in control, computing system design, communication, HD maps, testing, and human behaviors as one part of a three-article survey.

    Source: §I · pp. 1–2

  • Results

    The article provides a systematic, comprehensive survey of crucial AD and IV technology development with milestones, deployment details, and testing methods.

    Source: §I · p. 2

  • Takeaways & Limitations

    The combined work is intended to help abecedarians understand AD and IV development and bring insights to researchers as a bridge between past and future.

    Source: §I · p. 2, §VIII · p. 12

Abstract

from arXiv · show

Interest in autonomous driving (AD) and intelligent vehicles (IVs) is growing at a rapid pace due to the convenience, safety, and economic benefits. Although a number of surveys have reviewed research achievements in this field, they are still limited in specific tasks and lack systematic summaries and research directions in the future. Our work is divided into 3 independent articles and the first part is a Survey of Surveys (SoS) for total technologies of AD and IVs that involves the history, summarizes the milestones, and provides the perspectives, ethics, and future research directions. This is the second part (Part I for this technical survey) to review the development of control, computing system design, communication, High Definition map (HD map), testing, and human behaviors in IVs. In addition, the third part (Part II for this technical survey) is to review the perception and planning sections. The objective of this paper is to involve all the sections of AD, summarize the latest technical milestones, and guide abecedarians to quickly understand the development of AD and IVs. Combining the SoS and Part II, we anticipate that this work will bring novel and diverse insights to researchers and abecedarians, and serve as a bridge between past and future.

I. INTRODUCTION

This survey addresses the limits of task-specific AD and IV surveys by reviewing milestones across six technical areas, with this part covering control, computing system design, communication, HD maps, testing, and human behaviors. It also defines the field’s scope and organizes the chapters around task definitions, functional divisions, novel ideas, and milestones.

I. INTRODUCTION

  • The authors include L1–L5 vehicles in their working scope of AD, while noting that researchers disagree on which automation levels qualify.They distinguish automation-level taxonomies from the definition of AD and state that no unified definition exists.Source: §I · p. 2
  • The paper divides its technical review into six independent chapters, each covering task definitions, functional divisions, novel ideas, and AD/IV milestones.This part reviews control, computing system design, communication, HD maps, testing, and human behaviors; perception and planning are covered in Part II.Source: §I · pp. 1–2
  • The paper presents its survey as a systematic, comprehensive milestone review, with deployment details and testing methods across its technology sections.It is the second part of a three-article project intended to connect past and future work on AD and IVs.Source: §I · p. 2

II. CONTROL

  • The control section reviews vehicle-motion-control functions, control strategies, and state-of-the-art validation methods.It introduces the function of motion control and discusses both strategies and their validation.Source: §II · p. 2

A. Categories of Vehicle Motion Control

The section organizes vehicle motion control into longitudinal and lateral control, then surveys representative control methodologies and their trade-offs. Longitudinal control addresses speed and following behavior across traffic scenarios, while lateral control manages lane position and maneuvers.

A. Categories of Vehicle Motion Control

1) Longitudinal Vehicle Control:

  • Longitudinal control supports platooning, speed harmonization, trajectory smoothing, and signalized-arterial speed management, with string stability and communication-limited maneuvers among its design challenges.Strategies include fuzzy logic, PID, MPC, game theory, sliding mode, Lyapunov-based adaptive control, and AI approaches.Source: A. Categories of Vehicle Motion Control · p. 2

2) Lateral Vehicle Control:

  • Lateral control targets lane keeping, lane changes, and collision avoidance by minimizing lateral displacement and angular error, but weather, curves, and disturbances complicate real-world maneuvers.Reported approaches include game theory, PID, MPC, robust control, adaptive control, LPV controllers, sliding mode, fuzzy logic, and deep learning.Source: 2) Lateral Vehicle Control · p. 3

B. Control Methodology

1) PID Control:

  • PID control offers simple structure and implementation without detailed system knowledge, but parameter tuning is challenging and optimal performance is not guaranteed.The passage also notes difficulty adapting PID controllers to unknown and changing external environments.Source: 1) PID Control · p. 3

2) Game-theoretic Approaches:

The section surveys vehicle-control approaches, describing their mechanisms alongside the capabilities and constraints that shape their use.

2) Game-theoretic Approaches:

  • Game-theoretic control models vehicles as participants and traffic rules as part of decision-making, enabling interaction-aware choices such as lane changes.Accounting for mutual influence may increase decision reliability, while adding participants can raise computational complexity.Source: 2) Game-theoretic Approaches · p. 3

3) Fuzzy Logic Control:

  • Fuzzy logic control uses membership functions and if-then rules to map inputs to outputs without requiring a mathematical plant model.It can handle nonlinear vehicle dynamics, but parameter optimization is challenging and its inference rules require substantial expert knowledge.Source: 3) Fuzzy Logic Control · p. 3

4) MPC:

  • MPC repeatedly solves a predictive control problem and applies the first action sequence, balancing criteria such as comfort and fuel use within vehicle limits.Its handling of multiple criteria and physical and safety constraints is offset by difficulty with non-convex, high-complexity problems and real-time computation.Source: 4) MPC · p. 3

5) SMC Approaches:

  • SMC-based path-following controllers can adapt to unknown disturbances and matched uncertainties, with simulations reporting robust tracking performance.Time-varying disturbances and model inaccuracies in realistic scenarios can degrade control performance.Source: 5) SMC Approaches · p. 3

6) LPV Control:

  • LPV control is a linear approach used with predictive control for lateral and integrated-system control.It requires prior plant-model knowledge and real-time signals from all states; an observer is needed because those states are not provided by default.Source: 6) LPV Control · pp. 3–4

7) AI-guided Driving Policy Learning:

The section reviews learned driving policies and computing-system components, highlighting their capabilities and implementation constraints.

7) AI-guided Driving Policy Learning:

  • AI-guided systems can learn driving tasks within one framework and generalize to similar scenarios, but opaque behavior complicates safety validation and training-data needs may hinder research.Examples include online-learning road following, DQN lane changes, deep-reinforcement-learning braking, and CNN steering.Source: 7) AI-guided Driving Policy Learning · p. 4

A. Computing System Architecture

  • IV computing architectures include modular designs that separate computation, communication, storage, security and privacy, and power management, as well as end-to-end designs that map sensing data to control output.The survey focuses on modular architecture because it is described as more widely used and mature; modular separation can ease implementation, fault diagnosis, and updates.Source: A. Computing System Architecture · p. 4

1) Computing Hardware:

  • IV computing hardware includes CPUs, GPUs, FPGAs, DSPs, ASICs, and TPUs; the cited Xilinx FPGA achieved 14 images/W versus 4 images/W for the Nvidia Tesla K40 GPU.Source: 1) Computing Hardware · pp. 4–5

2) Operating System:

  • RTOS software manages resources for real-time autonomous perception, planning, decision-making, and control; QNX and VxWorks are common kernels.QNX includes CPU scheduling, interprocess communication, interrupt redirection, and timers; VxWorks is a monolithic kernel designed for real-time embedded systems and can also be used for safety and security.Source: 2) Operating System · p. 5

3) Middleware:

The section describes middleware’s role and computing-system evaluation metrics, then outlines constraints affecting IV performance, predictability, storage, thermal management, and power.

3) Middleware:

  • Middleware binds AD services and schedules application communication and sensing and computing resources, but can contribute to end-to-end latency and become a system bottleneck.Shared-memory communication was identified as an optimal solution for performance and reliability in a comparison study.Source: 3) Middleware · p. 5

4) Metrics for Computing Systems on IVs:

5) Constraints for Computing Systems on IVs:

  • Performance constraints include frame rate and processing latency; the cited safety target is vehicle response within 100 ms at a frequency of at least once every 100 ms.Predictability includes meeting time deadlines and making correct decisions, for which the 99.99th-percentile tail latency should be considered alongside mean latency.Source: 5) Constraints for Computing Systems… · p. 5
  • An IV can generate 2 to 40 TB of data per day, requiring substantial storage speed and space, while real-time storage and transfer increase energy and power use.Without cooling, each 1 kW of computing-system power can raise cabin temperature by up to 10°C per minute, motivating climate-controlled installation.Source: 5) Constraints for Computing Systems… · pp. 5–6
  • Computing, storage, and cooling power use can reduce vehicle mileage; each additional 400 W can reduce MPG by one, and GPUs can reduce fuel efficiency by up to 11.5%.The survey calls for more energy-efficient computing systems and green AI techniques.Source: 5) Constraints for Computing Systems… · p. 6

B. Sensor Systems

The section surveys onboard sensors and application-specific system design, alongside communication technologies that extend vehicles’ sensing and decision-making resources.

B. Sensor Systems

  • IV sensor systems include Lidar, Radar, cameras, ultrasonic sensors, IMUs, and GNSS; the survey emphasizes sensor fusion in response to driving uncertainty.Source: B. Sensor Systems · p. 6

1) Sensors in IVs:

  • Smart sensors provide extra detection or perception information, while non-smart sensors provide raw data for processing.The survey points to existing sensor-fusion methods and comparative studies of common onboard sensors.Source: 1) Sensors in IVs · p. 6

2) Computing System Design for IVs with Various Application Scenarios:

  • Computing-system design depends on application scenarios, policies, and regulations; mining trucks, passenger cars, and airport, harbor, and logistics vehicles have different requirements.Source: 2) Computing System Design for IVs… · p. 6
  • Explainable IV interfaces can provide onboard explanations of perception, localization, decision-making, and control using multimodal sensor signals.The survey distinguishes interpretability, including transparency and post-hoc explanation, from completeness, and notes that local and global explanations require different design choices [75].Source: 2) Computing System Design for IVs… · p. 6

IV. COMMUNICATION

  • V2X extends vehicle communication to vehicles, infrastructure, pedestrians, and clouds, using DSRC or C-V2X and supporting V2V, V2I, V2P, and V2C/N links.The section motivates dedicated communication because vehicle networks face rapidly changing topology, multipath and interference, and stringent latency and reliability needs for safety-related applications.Source: §IV · pp. 6–7

A. Standardization of V2X Communication

The survey outlines DSRC and C-V2X standardization, then surveys research directions for improving vehicular communication. C-V2X evolves from LTE-V2X to NR-V2X while retaining Uu and PC5 interfaces.

A. Standardization of V2X Communication

1) DSRC:

  • The U.S. DSRC example uses IEEE 802.11p and the IEEE 1609 family to form the WAVE protocol stack, while FCC allocations shifted spectrum toward C-V2X.The passage reports that the FCC allocated 30 MHz to C-V2X in 2020 and dropped support for DSRC.Source: 1) DSRC · p. 7

2) C-V2X:

  • C-V2X comprises LTE-V2X and NR-V2X phases, with NR-V2X extending 5G support for advanced services and adding unicast and multicast alongside broadcast.NR-V2X also enhances resource allocation and synchronization mechanisms.Source: 2) C-V2X · p. 8

B. Innovative Studies of Vehicular Communication

  • The survey frames communication research around physical-layer, media-access-control, routing, and security innovations for AD and IV requirements.The section introduces these directions as responses to stringent vehicular communication requirements.Source: B. Innovative Studies of Vehicular… · p. 8

1) Physical Layer:

The survey reviews communication techniques for reliable, efficient vehicular data exchange and introduces HD maps as a precise information source for AD functions.

1) Physical Layer:

  • OFDM is the major vehicular modulation technology, while millimeter-wave and visible-light communication are emerging directions for meeting AD data-rate needs.Research also optimizes OFDM parameters for reliable and efficient transmission.Source: 1) Physical Layer · p. 8

2) Congestion Control:

  • Limited bandwidth and massive terminal access can cause channel congestion and delayed or failed transmission; congestion-control approaches target stability and reliability.The passage lists power-, rate-, hybrid-, priority-, and cross-layer approaches.Source: 2) Congestion Control · p. 8

3) Resource Allocation:

  • Vehicular resource allocation optimizes channels, bandwidth, and transmit power, using centralized or distributed C-V2X schemes as well as learning-based methods.The passage also describes DSRC schemes for MAC parameters, channels, and rates.Source: 3) Resource Allocation · p. 8

4) Routing:

  • Routing supports long-distance exchange despite limited one-hop range and rapidly changing vehicle-network topology, with protocols optimizing QoS dimensions such as delay and reliability.The survey lists position-, topology-, geocast-, broadcast-, cluster-, and learning-based routing approaches.Source: 4) Routing · p. 8

5) Security:

  • Vehicular communication security research addresses attacks through authentication, authorization, confidentiality, integrity, and availability, with machine learning and blockchain as emerging directions.The passage names replay, man-in-the-middle, impersonation, spoofing, and malicious third-party attacks.Source: 5) Security · p. 8

V. HD MAP

  • HD maps support AD localization, perception, planning, control, and systems management, and can represent road elements at centimeter-level precision.Unlike commercial navigation maps, they serve the AD system rather than human drivers.Source: §V · pp. 8–9

1) Composition and Generation process:

The survey describes HD-map production, layering, standards, and quality assessment, then outlines vehicle-testing motivations and validation platforms.

1) Composition and Generation process:

  • HD-map production proceeds through sensor data acquisition, coarse map generation, semi-automated correction and verification, then testing, release, and updating.The map may be updated or removed when road elements or the region change.Source: 1) Composition and Generation process · p. 9

2) Types and Standards:

  • HD maps layer base, geometric, semantic, prior, and real-time information, while standards and industry forums support map storage, exchange, distribution, and generation.The survey names GDF, GDF-5.1, OADF, TPEG, ADASIS V3, NDS, and OpenDRIVE.Source: 2) Types and Standards · p. 9

3) Data quality:

  • Map quality is assessed through accuracy, precision, completeness, consistency, and timeliness, with cited work proposing geometry, localization-capability, and map-error assessment methods.Murphy and Pao [108] proposed detecting unmapped or mismapped roads and parking lots in map matching.Source: 3) Data quality · p. 9

VI. TESTING

  • Vehicle testing has drawn interest since 2016 amid accidents during testing or operation and growing interest in testing generalized intelligent systems.Fig. 4 shows a university test platform and an autonomous mining truck.Source: §VI · p. 9, Figure 4 · p. 9

A. Testing Platforms

  • Control validation uses simulation and experimental tests; simulations are less costly, faster, and more adaptable, while real-world testing costs time, labor, and money.The passage notes simulations can create scenarios difficult to reproduce in reality.Source: A. Testing Platforms · p. 9

1) Simulation Platforms:

The survey contrasts simulation and vehicle testing platforms, then organizes IV testing into function, performance, and intelligence tasks and discusses scenario-selection challenges.

1) Simulation Platforms:

  • Simulation platforms range from vehicle-dynamics tools to microscopic traffic simulators, but model errors require verification and validation because inaccurate models can make evaluation infeasible.Examples include Carsim, CARLA, PreScan, Matlab/Simulink, VISSIM 10.0, and AutonoVi-Sim.Source: 1) Simulation Platforms · pp. 9–10

2) Vehicles Platforms:

  • Small, lightweight test vehicles are rugged, low-cost, and portable but have limited payload, whereas on-road vehicles support validation across lighting and weather conditions and require approvals.The passage identifies ethical approval as an example of required approval.Source: 2) Vehicles Platforms · p. 10

B. Testing Methods

1) Testing Partition:

  • Function testing checks a specified output for a particular input and condition, performance testing covers input and condition ranges, and intelligence testing assesses strategies in broader scenarios.The three categories are function, performance, and intelligence testing.Source: 1) Testing Partition · p. 10

2) Testing Scenarios:

  • Scenario-based intelligence testing is difficult because necessary scenarios are numerous and expert-specified input-output testing is hard to conduct efficiently and reliably.ISO 26262 specifies functional outputs for scenarios, while ISO 21448 addresses functions affected by the external environment and foreseeable human error.Source: 2) Testing Scenarios · p. 10

3) Learning-based Approaches:

The section reviews testing challenges and human factors across automation levels, from driver assistance and takeovers to trust in highly automated vehicles.

3) Learning-based Approaches:

  • Difficult testing scenarios must be identified quickly while broad scenario coverage is pursued; balancing these goals remains unresolved, and finding corner cases merits further study [122].The review also describes virtual testing as a response to time-consuming real tests, with simulation focused on realistic behavior and varied scenarios [129].Source: 3) Learning-based Approaches · p. 10, 4) Parallel Testing · p. 10

4) Parallel Testing:

  • Many IV testing difficulties remain, and the field is forecast to enter a crucial development stage in the next decade.Virtual testing is prominent because real tests are time-consuming; simulated objects should exhibit rational behavior and varied manifestations.Source: 4) Parallel Testing · p. 10

VII. HUMAN BEHAVIORS IN INTELLIGENT VEHICLES

  • Human behavior matters to IV acceptance and mutual trust, but a unified framework for modeling human behavior still requires further investigation.The review organizes human factors by autonomy level because behaviors can differ significantly.Source: §VII · p. 10

A. Driver Assistance and Partial Driving Automation (L1-L2)

  • At L1–L2, drivers remain mentally in the control loop, motivating assistance research on fatigue, intention, attention, workload, emotion, and distraction.Methods are grouped as physiological-behavior-based or psychological-behavior-based approaches.Source: A. Driver Assistance and Partial… · pp. 10–11

B. Conditional Driving Automation (L3)

  • At L3, drivers may perform secondary tasks but must remain ready to take over, making time-constrained safe handovers a central human-factors concern.Research focuses on takeover algorithms, resuming control, and smooth transfers of control authority.Source: B. Conditional Driving Automation (L3) · p. 11

C. High/Fully Driving Automation (L4-L5)

  • With higher-level automation, drivers are no longer expected to maintain proper situation awareness to control the vehicle; trust largely depends on driving performance and is a major reason for AD/IV commercialization.An investigation of 162 Tesla drivers found that excellent introductions to system capabilities could establish high initial trust, and system usage increased over time regardless of automation failures.Source: C. High/Fully Driving Automation (L4-L5) · p. 11

D. Human-Machine-Interface (HMI)

The section reviews human-centered HMI design and intelligent cockpit challenges, then positions the survey as part of a three-part effort on AD and IVs.

D. Human-Machine-Interface (HMI)

  • AD HMI is framed as a human-centered system-engineering task for safety awareness, trust, comfort, and driver–vehicle collaboration.For L4/L5, the review covers five onboard HMI categories and design guidance for displays, time-constrained decisions, and driver monitoring.Source: D. Human-Machine-Interface (HMI) · p. 11

E. Intelligent Cockpit Systems

  • Intelligent cockpits must address safety, security, comfort, human factors, and energy use, beyond entertainment functions.The review notes that HMI has the greatest impact on comfortability among the listed factors, and that HMI can mitigate motion sickness.Source: E. Intelligent Cockpit Systems · pp. 11–12

VIII. CONCLUSION

  • The authors present this technical survey as a review of AD and IV research milestones, deployment details, and testing methods, intended to complement the other two parts.Together, the three articles are expected to offer insights and bridge past and future.Source: §VIII · p. 12

IX. BIOGRAPHY SECTION

The biography section profiles the contributors through their education, institutional roles, and research interests.

IX. BIOGRAPHY SECTION

  • Chen, Wang, Cao, Zheng, and Fei-Yue Wang are profiled with institutional roles and interests spanning autonomous driving, robotics, cognition, control, vision, and intelligent systems.Their biographies also describe academic appointments and professional service.Source: §IX · pp. 16–18
  • Li, Huang, Xing, Teng, and Lv are profiled with interests spanning computer vision, human-machine collaboration, driver behavior, parallel planning, and intelligent vehicles.Their biographies also report academic or research appointments and selected professional achievements.Source: §IX · pp. 16–17
  • Tian, Li, and Hu are profiled with interests in mobile computing, intelligent transportation, vehicular networks, human-machine interaction, computer vision, and deep learning.Their biographies describe university roles and related professional service.Source: §IX · p. 17
Loading 2305.11239v2…