You are on page 1of 17

1

A Survey of Simulators for Autonomous Driving:


Taxonomy, Challenges, and Evaluation Metrics
Yueyuan Li, Wei Yuan, Weihao Yan, Qiyuan Shen, Chunxiang Wang, and Ming Yang

Abstract—Simulators have irreplaceable importance for the the improvement of driving policy and route planning models
research and development of autonomous driving. Besides saving [3]. Moreover, simulators facilitate testing the accuracy and
resources, labor, and time, simulation is the only feasible way stability of control methods by simulating the dynamics behav-
arXiv:2311.11056v1 [cs.RO] 18 Nov 2023

to reproduce many severe accident scenarios. Despite their


widespread adoption across academia and industry, there is an ior of diverse vehicle types under various physics conditions
absence in the evolutionary trajectory of simulators and critical [4]. Additionally, simulators find applications in tasks such
discourse on their limitations. as vehicle cooperation and trajectory prediction related to
To bridge the gap in research, this paper conducts an in- autonomous driving [5].
depth review of simulators for autonomous driving. It delineates The simulators are widely adopted in academia and indus-
the three-decade development into three stages: specialized de-
velopment period, gap period, and comprehensive development, trial communities. In 2022, nearly half of the published meth-
from which it detects a trend of implementing comprehensive ods related to autonomous driving in IEEE Transactions on
functionalities and open-source accessibility. Then it classifies the Intelligent Transportation Systems were verified by simulators
simulators by functions, identifying five categories: traffic flow [6]. Leading technology companies in the autonomous driving
simulator, vehicle dynamics simulator, scenario editor, sensory industry, such as Waymo, Pony.ai, and Didi, emphasize the
data generator, and driving strategy validator. Simulators that
amalgamate diverse features are defined as comprehensive simu- development and application of simulators [7, 8, 9]. Waymo’s
lators. By investigating commercial and open-source simulators, impressive performance, as reflected in the Disengagement
this paper reveals that the critical issues faced by simulators Report from California between 2020 and 2023 [10], serves as
primarily revolve around fidelity and efficiency concerns. This compelling evidence of the critical role that realistic simulators
paper justifies that enhancing the realism of adverse weather play in achieving outstanding results [7].
simulation, automated map reconstruction, and interactive traffic
participants will bolster credibility. Concurrently, headless sim- A high-quality open-source simulator, which grants users
ulation and multiple-speed simulation techniques will exploit the access to the executable program and source code, holds
theoretic advantages. Moreover, this paper delves into potential significant advantages for both academic and industrial com-
solutions for the identified issues. It explores qualitative and munities. Firstly, it lowers barriers for individual researchers
quantitative evaluation metrics to assess the simulator’s per- and small institutions, who often find real-world testing unaf-
formance. This paper guides users to find suitable simulators
efficiently and provides instructive suggestions for developers to fordable and creating a simulation platform time-consuming.
improve simulator efficacy purposefully. With a free and open-source simulator, more researchers are
encouraged to engage in the development of autonomous
Index Terms—Autonomous Driving, Simulator, Testing.
driving technologies, fostering innovation and collaboration
within the field. Secondly, an ideal open-source simulator
I. I NTRODUCTION serves as a valuable platform for data sharing. This enables
Simulators play a crucial role in the field of autonomous researchers to exchange data related to vehicles, sensors,
driving. According to Kalra et al., validating the performance and traffic scenarios, leading to substantial time and cost
of an autonomous vehicle with statistical significance would savings in data collection efforts. Additionally, the diversity
require an impractical 215 billion miles of real-world testing of shared data contributes to comprehensive improvements in
[1]. With simulators, this process can speed up with lower autonomous driving systems. Furthermore, data sharing within
resource costs. Simulators offer the ability to generate raw the simulator ecosystem can promote the establishment of
sensor data and ground truth under adverse weather condi- standards in data format and communication protocols. Rec-
tions, which helps to enhance the robustness of perception ognizing the importance of public simulators, influential insti-
algorithms [2]. They also enable the simulation of complex tutions like Computer Vision Center (CVC) and Baidu have
traffic scenarios, including severe accidents, which aids in developed simulators such as CARLA and Apollo [11, 12].
These initiatives reflect the growing acknowledgment of the
This work is supported by the National Natural Science Foundation of value of open-source simulators for autonomous driving.
China (62173228). Ming Yang is the corresponding author.
Yueyuan Li, Weihao Yan, Qiyuan Shen, Chunxiang Wang, and Ming Yang The history of traffic flow simulators serves as a valuable
are with the Department of Automation, Shanghai Jiao Tong University, reference, highlighting the evident shift towards open-source
Key Laboratory of System Control and Information Processing, Ministry of simulators as the future trend. Initially, most traffic flow sim-
Education of China, Shanghai, 200240, CN (email: MingYANG@sjtu.edu.cn).
Wei Yuan is with the Innovation Center of Intelligent Connected Vehicles, ulators, such as Paramics and Aimsun, were commercial soft-
Global Institute of Future Technology, Shanghai Jiao Tong University, Shang- ware [13, 14]. However, with the decreasing cost of software
hai, 200240, CN development, the open-source simulator SUMO emerged and
steadily gained popularity, ultimately becoming the preferred
2

choice for traffic flow simulations [15], while Paramics and began to emerge. Zhou et al. conducted a detailed survey
Aimsun faced the challenge of acquisition. Advancements in of seven popular simulators, including CARLA, LGSVL,
key techniques required for simulator development, such as CarSim, AirSim, Prescan, Matlab/Simulink, and CarMaker.
3D modeling, graphical rendering, and parallel computing, Their work explored the functionalities of these simulators and
have become more accessible and affordable. The cost of summarized their respective application scopes [28]. Holen et
developing simulators is expected to decrease, leading to a al. conducted a similar study but expanded the range of simu-
decline in simulator license prices. As a result, users will lators, offering a broader perspective on the available options
naturally gravitate towards more economical choices, namely, [29]. The simulators mentioned in Kaur et al.’s research were
open-source simulators. mostly consistent with Zhou’s work, with the main difference
This paper serves as a comprehensive guide for researchers being the replacement of AirSim with Gazebo in their survey.
in the field of autonomous driving, assisting them in selecting This survey qualitatively describes the performance of the
suitable simulators for their specific tasks. It emphasizes simulators, along with a summary of the desired functionalities
the current limitations of simulators, alerting researchers and and characteristics of simulators [30]. The existing reviews
developers of the potential factors that may defect the validity tended to assist researchers to choose simulators, so their
of simulation-based experiments. This paper further offers main content is to list the available alternatives and check the
insightful suggestions to the developers of open-source sim- detailed functions. There remains a need for more comprehen-
ulators, directing their attention toward critical issues worthy sive and in-depth reviews to discuss simulators’ performance,
of addressing. The contribution of this paper is fourfold: limitations, and potential directions for improvement.
1) It presents a comprehensive overview of the evolution Over the past decade, discussions regarding critical issues
of simulators over the past thirty years and predicts the in simulators have been scattered, with various researchers
likely development tendencies in the future. highlighting specific challenges. Hu et al. stressed the need
2) It categorizes the simulators based on their utility. Spe- to narrow the gap between simulators and the real world
cial attention is given to identifying free and open-source by automating the reconstruction of simulation scenarios and
simulators for potential users. improving the sim2real model transfer efficiency [31]. Mean-
3) It identifies and defines the critical issues in current while, Zhang et al. identified that enhancing the intelligence of
simulators from a developer’s perspective. It further in- traffic participants’ behaviors is a promising avenue to boost
vestigates the potential solutions to address these issues. simulator performance [3]. Similarly, Siebke et al. emphasized
4) It reviews the evaluation metrics to measure the effi- the importance of including human errors in the traffic sim-
ciency and fidelity of a simulator, which is organized ulation process to enhance realism [32]. While these papers
by the mechanism of algorithms. succeeded in justifying the importance and urgency of these
issues, they often fell short in relating the existing technology
II. R ELATED W ORKS to potential solutions and providing compelling evaluation
The limited number of reviews on simulators for au- metrics to measure the effectiveness of proposed resolutions.
tonomous driving is primarily due to the lack of simulators. Consequently, the development community has not agreed
Most surveys and reviews in the field of autonomous driving on the simulators’ insufficiency, leading to unorganized code
have focused on other aspects, such as tasks, algorithms, development and aimless release of new simulators.
and sensors. Though these reviews briefly mention the use This paper aims to address this gap in the review of
of simulators in various scenarios, they often do not exten- simulators. In addition to investigating simulators’ history,
sively cover the topic [16, 17, 18]. For example, Yurtsever category, and available items, this paper will analyze and
et al. demonstrated the capability of simulators to generate connect existing technologies to viable resolutions, providing
diverse data that captures different weather conditions [19]. a clear roadmap for developers to focus their efforts and
This data augmentation method enhances the robustness of create more effective and efficient simulators. By proposing
perception models, enabling them to perform effectively in well-defined evaluation metrics, this paper seeks to establish
a wide range of scenarios [20]. Similarly, Grigorescu et al. a standardized way to measure the performance and quality
emphasized the importance of simulation for deep learning- of simulators. By filling this void in the literature, this paper
based driving decision-making models, as these models require benefits researchers, developers, and the autonomous driving
diverse interaction behavior to improve their decision-making community as a whole.
capabilities [21]. This viewpoint is echoed in other reviews that
focus on decision-making algorithms [22, 23, 24, 25]. What’s III. P ROBLEM R EFINEMENT
more, Tampuu et al. expressed concerns about the fidelity of Defining the scope of discussion is crucial to ensuring a
simulators, highlighting the challenges of transferring decision clear and focused review. In the context of this paper, the
models trained in virtual environments to the real world term simulator refers to software tools that replicate the real
[26]. While these reviews provide strong evidence of the world with sufficient detail for developing functions within
widespread application of simulators across various tasks, it is autonomous driving systems. The scope excludes simulators
worth noting that most of these reviews merely list alternative integrated with hardware that aim to mimic a realistic driving
simulators without delving into their quality [27]. experience by technology like digital twins. Considering the
It was not until 2020, with the increasing number of public substantial engineering effort required to develop usable sim-
simulators, that reviews and discussions focusing on simulators ulators, the number of mature and readily available options is
3

limited. Hence, this review will not be restricted to comprehen- Additionally, software techniques such as rendering, object
sive simulators capable of supporting multiple research topics modeling, and physics engine were under development. Con-
across various areas such as perception, planning, and control. sequently, constructing a high-fidelity simulator was daunting
Instead, it will encompass various specific-purpose simulators. and costly. During that period, the academic community had
Since commercial simulators constitute a significant portion not recognized the potential of learning-based methods, so
of available options and hold a dominant position in some there was no urgent demand for generating a large volume
categories, they will be included in the review of history and of high-fidelity data. The combination of technological limi-
taxonomy. However, due to commercial copyright restrictions, tations and absent motivations led to a ten-year gap. Only a
performance evaluations of these commercial simulators will few simulators managed to survive and make a lasting impact
not be provided in this paper. Instead, the focus will be on during the harsh period. Notable among them were PreScan,
free and open-source simulators when discussing evaluation supported by Siemens in terms of funding and hardware [38];
metrics, performance, and existing issues. rFpro, which benefited from the financial backing and data
support of Formula 1 (F1) racing groups [37]; and VI-grade,
a comprehensive investment by HBK Investment [39]. These
IV. H ISTORY OF S IMULATORS
simulators’ survival relied heavily on substantial investments
The history of simulators for autonomous driving provides from world-leading profit organizations.
valuable insights into the technological evolution within this The resurgence of simulators began in 2013, and the focus
domain. Over the past three decades, the development of sim- evolved from singular functionalities to the diverse needs of
ulators has progressed through distinct stages: the specialized multiple tasks. One influential factor was the release of the
development period, the gap period, and the comprehensive video game GTA V in 2013, which served as an effective
development period. Two noteworthy trends have emerged tool for data generation owing to its advanced rendering
in the development of simulators. Firstly, there is a notable technology. Researchers leveraged this platform to collect
expansion of simulator capabilities, enabling the integration of images under various weather conditions at a low cost [40].
multiple functions on a single platform. This versatility allows The expansive city map of GTA V also attracted researchers
for the efficient development of diverse tasks within the same to employ it in driving decision-making models [41]. Despite
simulator environment. Secondly, the idea of free and open- Take-Two Interactive, the developers of GTA V, rejecting
source simulators has grown popular. An increasing number the autonomous driving interface in 2016, the application
of companies and organizations are contributing open-source demonstrated the immense potential of simulators. It showed
software to establish their influence on evaluation benchmarks. the advantages of simulators to replicate diverse weather
The period before 2002 is described as the specialized conditions, offering researchers the flexibility to study various
development period due to the focus of simulators on singular scenarios. Some subsequent 3D simulators emerged, aiming to
objectives. Most of the simulators worked on two topics, traffic approximate the real world and provide diverse synthetic sen-
flow, and vehicle dynamics, respectively. Traffic flow simu- sor data, actively offering interfaces to access proxy vehicles
lation proved instrumental in road planning and optimizing for validating driving decision algorithms [11, 42, 43]. In 2018,
traffic efficiency, while vehicle dynamics simulation facilitated Waymo utilized bird-eye view (BEV) semantic images to train
the assessment of component adjustments on vehicle physics driving decision models, which demonstrated an impressive
behavior. The maturity of relevant theories and technologies performance in trajectory prediction with an easily accessible
during this era significantly contributed to the emergence of input [44]. This approach encouraged the development of
traffic flow and vehicle dynamics simulators. These simulators 2D BEV simulators such as Highway-env and CommonRoad,
were built upon explicit rules and physics models, requiring tailored to specific simulation requirements and research ob-
relatively modest computational resources. In 1992, PTV jectives [45, 46]. These developments marked a significant
Group introduced PTV Vissim for traffic flow and logistics shift in the applications of simulators for autonomous driving.
system simulations [33]. Later, in 1994, Paramics appeared to Besides the shift in functionality, the history of simulators
offer similar functionalities [13]. Alongside these commercial also reflects a trend toward open-source. Fig. 1 illustrates
options, the open-source simulator SUMO played a critical the changes in the number of open-source and closed-source
role in traffic flow simulation [34]. Concurrently, in the realm simulators over the past three decades. The data clearly
of vehicle dynamics, CarSim and IPG CarMaker, established reveals a consistent growth of free and open-source simulators.
during the 1990s, fostered enduring partnerships with nu- This trend can be attributed to two main factors. Firstly, the
merous vehicle manufacturers [35, 36]. Their prominence in technologies required for developing simulators have become
this domain was reinforced by the extensive real-world data more accessible over time. The decreasing costs and increasing
they accumulated. RFpro, which appeared a decade later in availability of computing resources and software tools have
2007, was one the few simulators boasting a vehicle dynamics made it easier for developers to create and distribute open-
fidelity comparable to CarSim and IPG CarMaker, owing to source simulators. Therefore, more researchers and devel-
the real-world data collection from racing cars [37]. opers are contributing to open-source simulators. Secondly,
From 2003 to 2013, the development of simulators faced the emergence of open-source simulators is beneficial for
significant challenges as related technologies encountered bot- the entire autonomous driving industry. These platforms can
tlenecks. Primarily, computer hardware, especially graphics promote the establishment of standardized data formats and
cards, lacked the capability to handle large-scale rendering. evaluation criteria, and further foster a more cohesive and
4

Fig. 1. The number of simulators growing by year. This statistical result includes the popular commercial simulators used by famous vehicle manufacturers
or technology companies and open-source simulators with over 100 stars on GitHub or more than 100 citations on paper.

efficient research environment, where findings can be better influence while indicating their current status, such simulators
shared and validated. These two advantages together propel are denoted with gray text.
the growth of open-source simulators.

V. C ATEGORIZATION OF S IMULATORS A. Traffic Flow Simulator

The development history of simulators reveals their origin Traffic flow simulators play a crucial role in simulating
from integrated programs designed to automate the training and analyzing the behavior of vehicles and pedestrians in
and testing processes for various functions within autonomous a traffic system. Their primary objective is to assess and
driving systems. Therefore, it is natural to categorize the sim- optimize traffic systems’ performance, safety, and efficiency,
ulators based on their functionalities. Employing this frame- providing valuable insights for engineering tasks such as urban
work, simulators are classified into five specialized classes: planning and road arrangement. These simulators commonly
traffic flow simulator, vehicle dynamics simulator, scene editor, offer functionalities to edit the road network, traffic signs, and
sensory data generator, and driving strategy validator. When signal timing and adjust the behavior patterns of vehicles and
a simulator can handle multiple tasks, it is positioned within pedestrians, including factors like average vehicle speed, den-
the sixth category, a comprehensive simulator. sity, and lane-changing frequency. Simulation outputs typically
TABLE I demonstrates statistics on the influential sim- encompass metrics such as congestion, emissions, and traffic
ulators. The table includes essential information about the capacity. The focus of these simulators is on macroscopic
simulator’s name, type, first release date, last maintenance aspects of traffic conditions rather than individual traffic
time, developer, accessibility, and supported systems. The rows participant movement trajectories, making high accuracy in
are sorted according to the simulators’ categories, focusing on vehicle dynamics and map details unnecessary.
their most representative features, even as they may expand to Since the 1990s, various traffic flow simulators like PTV
include new functionalities, which ensures clarity and avoids Vissim, Aimsun, Paramics, and SUMO have been extensively
ambiguity in the classification process. The statistical results used [13, 14, 15, 33]. In recent years, there is no new simu-
consider both commercial and open-source simulators, filtered lators for traffic flow simulation because the functionalities of
by their performance and influence. This table includes com- existing ones have been stabilized. Notably, the open-source
mercial simulators used by renowned vehicle manufacturers simulator SUMO continues to update its microscopic vehicle
or technology companies, as well as free and open-source models. Its open-source nature has garnered support from the
simulators that have gained substantial attention, with over community, with many developers connecting their driving
100 stars on GitHub or more than 100 citations on Google strategy validators to SUMO to benefit from its ability to
Scholar. Especially, some simulators had a significant impact control the background vehicles in the virtual environment
but are no longer actively maintained. To acknowledge their [61].
5

TABLE I
BASIC INFORMATION OF CATEGORIZED SIMULATORS .
I F THE SIMULATOR IS NO LONGER UNDER MAINTENANCE , THE TEXT WILL BE COLORED GRAY.

Open-source Supported systems


Simulator Category Maintenance Provider
and free Windows Linux

PTV Vissim [33] Traffic flow 1992-now PTV Group √ √
Paramics [13] Traffic flow 1994-now SYSTRA √ √
Aimsun [14] Traffic flow 1999-now Siemens √ √ √
SUMO [15] Traffic flow 2001-now Eclipse Foundation
√ √
CarSim [35] Vehicle dynamics 1996-now Mechanical Simulation √ √ √
Webots [47] Vehicle dynamics 1998-now Cyberbotics √ √
CarMaker [36] Vehicle dynamics 1999-now IPG Automotive √ √ √
Gazebo [48] Vehicle dynamics 2002-now Open Robotics √ √
rFpro [37] Vehicle dynamics 2007-now rFpro
√ √
VI-Road [39] Scenario editor 2007-now VI-grade √ √
RoadRunner [49] Scenario editor 2020-now MathWorks
√ √
Parallel Domain [50] Sensory data 2017-now Parallel Domain √ √ √
AirSim [42] Sensory data 2017-2022 Microsoft √ √ √
SVL [43] Sensory data 2018-2021 LG Electronics √ √
VI-WorldSim [39] Sensory data 2020-now VI-grade √ √
Nvidia Drive Sim [51] Sensory data 2020-now Nvidia √ √ √
Vista [52] Sensory data 2020-2022 Vista Group
√ √ √
Torcs [53] Driving strategy 1997-2020 Torcs Group √ √ √
CarRacing [54] Driving strategy 2014-now Farama Foundation √ √ √
VDrift [55] Driving strategy 2014-now VDrift Group √ √ √
Udacity Self-driving [56] Driving strategy 2016-2022 Udacity √ √
Simulink Driving strategy 2017-now Mathworks √ √
Apollo [12] Driving strategy 2017-now Baidu √ √ √
CommonRoad [46] Driving strategy (2D) 2017-now TUM √ √ √
highway-env [45] Driving strategy (2D) 2018-now Eduoard Leurent √ √
SMARTS [57] Driving strategy (2D) 2020-now Huawei Noah √ √ √
SUMMIT [58] Driving strategy 2020-2022 AdaCompNUS
√ √
SCANeR Studio [59] Comprehensive 1990-now AVSimulation √ √
Virtual Test Drive [60] Comprehensive 1998-now Hexagon √
Prescan [38] Comprehensive 2006-now Siemens √ √ √
CARLA [11] Comprehensive 2016-now Computer Vision Center

B. Vehicle Dynamics Simulator been developed to provide a realistic dynamic system for
racing cars and is widely utilized by F1 teams [37]. Among
Vehicle dynamics simulators are specifically designed to
open-source options, Gazebo and Webots are reliable choices
simulate the dynamic behavior of vehicles using principles
due to their rich experience in physics simulation in robotics
of physics. These simulators take into account both internal
[47, 48]. These two simulators keep improving their vehicle
factors, such as the vehicle’s powertrain, suspension, and
dynamics simulations to meet the growing demands of users.
aerodynamics, and external conditions, such as road friction,
wind resistance, and slope. Considering these variables, the
simulator can accurately estimate the dynamics characteristics C. Scenario Editor
of vehicles, including acceleration, braking, and steering re- A scene editor provides a visual interface for editing maps
sponse to different control commands. Their widespread use and trajectories of traffic participants. The editing of maps
includes evaluating vehicle performance, optimizing vehicle involves modifying the geometric properties and topological
design, and enhancing safety performance. relationships of roads, as well as adding points of interest
Due to the well-established theories in vehicle dynamics, such as traffic lights and road signs. On the other hand, editing
the corresponding simulators have been available for a con- traffic participants entails adjusting their appearance, type, and
siderable period. However, because of the strong reliance trajectories within the scene. As is demonstrated in Fig. 2, the
on real-world experimental data, the most reliable vehicle purpose of scene editors is to provide data for downstream
dynamics simulators are typically developed for commercial tasks like traffic flow simulation, sensor data generation, and
use in close collaboration with car manufacturers. Leading the validation of driving strategies, where the edited data is
examples of state-of-the-art (SOTA) simulators in vehicle further processed and utilized. Therefore, scene editors are
dynamics are CarSim and IPG CarMaker, famous for their typically integrated as submodules within other simulators
outstanding performance [35, 36]. For vehicle manufacturers rather than existing as standalone tools.
seeking highly accurate estimation of vehicles’ physical behav- Among the popular scene editors, the 2D scene editor
ior, collaborating with these simulators is often the preferable attached to SUMO stands out as an effective and widely used
choice. In the domain of F1 racing, the rFpro simulator has tool [15]. For 3D scene editing, MathWorks’ RoadRunner
6

Fig. 2. The common tasks in developing an autonomous driving system and their corresponding simulator types [11, 15, 34, 35, 42, 46, 49]. If a simulator
can cover all the tasks within the system, it can be regarded as a comprehensive simulator.

is a notable choice and distinguishes itself as a standalone rendering typically requires game engines, many perception
simulator [49]. It is essential to note the significant difference data generators offer interfaces for controlling agent vehi-
in how maps and vehicle trajectories are represented between cles. The distinction between a perception data generator
2D and 3D scene editors. Consequently, manual adjustments and a comprehensive simulator lies in vehicles’ physical
may be required when converting data formats [34, 62]. The performance and traffic participants’ behavior patterns. These
issue of the map’s data format will be further discussed in features determine whether a simulator can serve functions
detail in Section VI-B. beyond data synthesis. Accordingly, simulators like LGSVL
and AirSim are classified as perception data generators, given
D. Sensory Data Generator their primary focus on generating sensor data rather than
providing a comprehensive simulation environment [42, 43].
Sensory data generators are tools that can synthesize raw
sensor data and ground truth annotations using rendering and
numerical calculation techniques. The sensor data typically
E. Driving Strategy Validator
includes RGB images, lidar points, and radar points, while
ground truth annotations contain objects’ bounding boxes, Driving strategy validators examine the performance of
semantic labels, and depth values. These generators play a algorithms in tasks such as path planning, trajectory prediction,
crucial role in improving the robustness of perception al- and driving policy generation by providing executable traffic
gorithms, as most SOTA algorithms are data-driven. Since scenarios. A branch of driving strategy validators emphasizes
collecting real-world data under diverse conditions is costly decision-making as a submodule of an autonomous driving
and time-consuming, researchers have to depend on sensory system, thus pursuing a high fidelity in sensor data and
data generators to enhance the diversity and complexity of vehicle dynamics. The main difference between this branch
their datasets [1]. Despite the significance of these generators, of driving strategy validators and comprehensive simulators
their numbers remain limited due to the ongoing development lies in scene editing. Due to the high cost and complexity
of synthesizing algorithms [63, 64, 65]. Parallel Domain is one of obtaining map and object trajectory data in a scene, most
of the few companies that have successfully commercialized driving strategy validators prefer to provide these data as part
this data synthesis service [50]. of their paid services without giving users sufficient freedom
Perception data synthesis is often conducted through the in editing. Examples of simulators in this category include
rendering functionality within simulators. Since realistic scene Apollo, Udacity self-driving platform, and Vista [12, 52].
7

Fig. 3. The critical issues in simulators [63, 66, 67]. The fidelity problems affect the reliability of the simulation. The efficiency problems affect the time
and computation resource cost of the simulation.

Since high-quality validators are usually commercial soft- traffic participants. A comprehensive simulator of high quality
ware or paid services with high hardware requirements, they can notably accelerate the research of autonomous driving.
are unaffordable for many researchers. This situation motivates Nevertheless, the development process requires multidisci-
the development of small and independent driving strategy plinary technologies and substantial financial support.
validators, which use simplified vehicle dynamics models and The Virtual Test Drive (VTD) is a widely acclaimed com-
idealized perception data. These validators focus on verifying prehensive simulator, offering highly integrated functionalities
the effectiveness of algorithms while underlooking the errors that satisfy diverse needs in developing an autonomous driving
brought by sensors and vehicle control. In the open-source system [60]. A similar product, SCAnER, presents comparable
community, this branch of simulators has gained widespread attributes, yet its popularity remains relatively limited due
attention. Two good examples include TorcsRL, which focuses to a lack of promotion [59]. Waymo is reputed to possess
on validating reinforcement learning-based control models an advanced comprehensive simulator, though the technical
in racing scenarios [53], and Highway-env, which provides intricacies have not been disclosed publicly [7]. Within the
four environments (highway, intersection, roundabout, and realm of open-source simulators, CARLA stands out as the
parking lot) for evaluating the performance of driving policy sole platform demonstrating the inclination and capability to
models [45]. These open-source simulators provide convenient evolve into a comprehensive simulator. Still, there is much
and flexible tools that allow researchers to quickly validate space for improving its performance [11].
their theories and algorithms without worrying about complex Despite the challenges in developing a comprehensive sim-
dynamics models and expensive perception data collection. ulator, it is meaningful for autonomous driving. Such a simu-
lator can present realistic scenarios and situations, allowing
F. Comprehensive Simulator researchers and developers to assess the performance and
safety of autonomous driving systems thoroughly. Therefore,
Simulators mentioned above are tailored for specific tasks.
continued investment and support for developing simulators
However, variations in data format between the simulators
are crucial for advancing autonomous driving technology.
from different developers can lead to errors and even sys-
tem crashes during testing. Comprehensive simulators offer a
unified environment that promotes seamless interaction among VI. C RITICAL I SSUES
components, resolving format inconsistencies and ensuring a Simulators for autonomous driving face challenges in fi-
smoother testing process for the entire system. Fig. 2 shows delity and efficiency (see Fig. 3). Fidelity in simulation per-
the tasks in autonomous driving systems that can be fully tains to the discernible variance between synthesized data and
developed within a virtual environment and their relation with real-world data. In tasks including adverse weather conditions,
the task-specific simulators. Comprehensive simulators stand automated map reconstruction, and interactive traffic partici-
out as entities that integrate a majority of functionalities, which pants, the output produced by simulators exhibits significant
endows them with the capacity to handle relatively complete disparities from the real-world data [31, 32, 68]. Efficiency
autonomous driving systems. These simulators encompass a in simulation centers around code optimization, thereby ex-
traffic scenario editor, generate highly realistic perception data, pediting the training and testing phases. Headless simulation
and provide interfaces for program-controlled vehicles. In the enables the deployment of simulators across large-scale, re-
simulation process, comprehensive simulators consider the mote servers without a physical display, which makes better
vehicles’ physics behavior and interactions with surrounding use of computation resources [69]. High frame-rate simulation
8

underscores the refinement of simulation algorithms, culminat-


ing in accelerated individual simulation processes [70, 71, 72].
Addressing these challenges is crucial as current simulators
often suffer from distortion, undermining the credibility of
simulating experiments; the high hardware requirements and
limited simulation speed indicate that simulators have not fully
realized their theoretical advantages.
By observing the free and open-source simulators, this
section delves into the critical issues in simulators, providing
definitions and emphasizing their importance. It will also
review existing methods to tackle these problems, offering
potential solutions for enhancing simulators’ performance.

A. Adverse Weather Simulation


The common sensors in autonomous driving systems in-
clude the camera, lidar, radar, IMU, speedometer, and GNSS.
Among these, the camera, lidar, and radar take center stage
in local perception tasks, serving to dynamically comprehend
the traffic environment encompassing an autonomous vehicle.
Regrettably, these cameras dedicated to real-time local percep-
tion are notably vulnerable to adverse weather conditions [73].
Factors like illumination, visibility, and humidity fluctuate in Fig. 4. Particle-based snowfall simulation for lidar points [65]. From the
different weather, introducing noise into sensor data, so the comparison, the augmented data improve the accuracy of object detection,
precision and range of perception outcomes are compromised
[74]. Adverse weather conditions also alter road reflectance,
further perturbing the efficacy of perception algorithms [75].
Hence, improving perception algorithms’ robustness in diverse
weather has been a paramount objective for a long time.
Collecting authentic sensor data from real-world scenar-
ios is costly and time-demanding. Given the erratic nature
of weather, collected data might deviate unpredictably from
anticipated norms. In contrast, simulators offer precise control
over weather conditions through adjustable parameters. They
circumvent the need for prolonged waits to encounter specific
weather scenarios [1]. Moreover, the labor-intensive task of
manually annotating ground truth data is obviated. These
inherent advantages suggest that if simulators can proficiently
generate high-fidelity data in adverse weather conditions, they
possess the potential to supplant real-world data and promote
relevant research [74, 76].
Most existing simulators focus on the presence of ad-
verse weather simulation functionality without considering the
actual quality of the implementation. Since the simulators
usually rely on game engines for development, the common Fig. 5. GAN-based rain simulation for RGB images[64]. Tremblay et al. apply
choice is to use the engines’ default particle system to generate a GAN-based method to simulate a rainy scenario with 50 mm/hr precipitation.
adverse weather [11, 42, 43]. In a particle system, rain and
snow are treated as particles with physics properties. During
the simulation process, the emphasis is on the movement tra- the distribution, and the original data merge with the mask
jectory of particles due to wind speed and gravity, but less at- through post-processing. Modeling the noise distribution pat-
tention is paid to the texture templates of particles [63, 64, 77]. tern with particle systems and physical properties was one
Thus the image data generated based on particle systems lacks of the earliest practices [63, 78, 79], which was initially
realism. Furthermore, existing weather simulations often only applied to image generation and extended to enhance lidar
affect rendering, meaning they can only generate image data point data (see Fig. 4) [65]. In recent years, learning-based
for adverse weather conditions without adding corresponding data synthesis methods have been increasing [80, 81]. Among
noise interference to lidar and millimeter-wave radar. them, approaches based on Generative Adversarial Networks
In most synthesis processes, the distribution pattern of (GANs) and Neural Radiance Fields (NeRF) have achieved
noise is first extracted, then a mask is generated based on impressive results [82, 83, 84, 85]. Fig. 5 illustrates rain
9

data synthesized by Tremblay et al. using a combination


of CycleGAN and physics-based rendering. This synthesis
maintains visual realism while adhering well to physical laws.
In order to enhance simulators’ performance in simulat-
ing adverse weather conditions, the simulator should pro-
vide a flexible and user-friendly interface that allows users
to integrate customized data synthesis models. Advanced
weather simulation algorithms are often available as open-
source projects online [76, 81, 85]. These algorithms vary in
environment configuration and hardware requirements, posing
challenges for users looking to adopt them quickly. A solution
could be establishing a standardized parameter interface within
the simulator, allowing users to effectively adjust various
weather parameters to simulate different weather conditions. Fig. 6. Surface reconstruction based on lidar points [92]. This scene is
reconstructed based on data from Waymo’s dataset, which consists of 10M
points, taking only 20s.
B. Automated Map Reconstruction
High-definition maps are pivotal in crafting the testing traffic scenario, the 2D maps generated by these approaches
environments for autonomous driving systems. A meticulously become preferable to the development of driving decision-
annotated map is the foundation for generating sensory data making algorithms [68, 96].
and formulating driving strategies. Given that the intricate These two categories of automated map reconstruction
road layouts and exceptional vehicle behaviors predominantly methods can reduce the need for manual intervention in
originate from real-world scenarios, generating traffic scenar- constructing simulation scenarios for autonomous driving [97].
ios, especially the maps, hinges heavily on authentic data. However, they are unsuitable for practical implementation at
The conventional map reconstruction process manually ex- current. The 3D map construction techniques exhibit sensitiv-
tracts road structures, labels road network relationships, and ity to data distribution and susceptibility to noise interference
annotates traffic signals from lidar points and images, which is [66]. The precision and speed of these methods fall short
labor-consuming. By comparison, automated map reconstruc- of the standard for industrial application [93]. It is worth
tion utilizes algorithms to process 2D pixels and 3D points noting that the output of 3D map reconstruction is the 3D
and generate structured maps. Achieving partial or complete mesh model, which lacks structural information on roads and
automated map construction can significantly reduce manual buildings. Also, semantic annotations about signs and signals
effort and improve the efficiency of building scenarios. remain absent within the 3D mesh model. Compared to the
Automated map reconstruction is advancing along two widely adopted 3D map data format OpenDRIVE, 3D mesh
distinct trajectories. One avenue is oriented towards 3D re- models exhibit significant data gaps [62]. Although 2D map
construction, characterized by an emphasis on achieving high- reconstruction approaches furnish data on road structure, the
level realism and details. The general idea is to extract features seamless integration of their outputs with 3D map reconstruc-
from 2D images and match them with 3D points and pose tion poses challenges. Typically, the outcomes of 2D map
information. By establishing correlations between object mod- reconstruction are expressed in XML-based formats like Open-
els and textures, this method enables effective automatic map StreetMap (OSM) and Lanelet [98, 99]. Due to the different
reconstruction [86, 87, 88]. Traditional 3D map reconstruction mechanisms by which OSM and Lanelet store road geometry
methods use explicit priors to process large-scale raw data compared to the OpenDRIVE, data conversion often raises
rapidly [88, 89]. The new rising learning-based methods ex- errors necessitating manual correction. Therefore, matching
tract data-driven priors by neural networks to handle noise and the road structure with the 3D mesh model is a task that
data sparsity [90, 91]. Fig. 6 depicts the reconstruction result remains accomplished. Addressing these challenges is critical
of an advanced 3D map reconstruction method, Neural Kernel for advancing the state of automated map reconstruction and
Surface Reconstruction [92]. A recent breakthrough in this realizing its potential in autonomous driving simulation.
area is the NeRF algorithm. It learns the implicit features of
space from images captured from different angles, representing
3D space as a continuous radiance field [93]. NVIDIA has C. Interactive Traffic Participants
implemented 3D surface reconstruction and rendering based A prominent limitation of current simulators is the absence
on the NeRF algorithm, demonstrating its great potential of intelligence in traffic participants. Traffic participants en-
for automated 3D map reconstruction [67]. Another avenue compass vehicles, cyclists, pedestrians, and dynamic objects
focuses on reconstructing the road structure and relationships, within a traffic scenario surrounding the self-driving vehicle.
exhibiting a greater tolerance for deviation in map details Ideally, these participants should be able to perceive their
[94]. These techniques prioritize comprehensively depicting surroundings, comprehend the actions of other participants,
the traffic environment, encompassing elements such as the and navigate within the scenario efficiently [95].
road network, traffic signals, and dynamic objects [66, 95]. Presently, the traffic participants featured in simulators rely
With the overarching structure and interplay dynamics of the on rule-based behavioral models, which exhibit simplistic
10

D. Headless Simulation
Headless simulation empowers a simulator to operate on
remote servers without a physical display. By eliminating the
necessity for a Graphical User Interface (GUI), the simula-
tor can allocate computational resources more streamlined,
leading to optimal resource utilization. This approach proves
particularly beneficial for the seamless deployment of simula-
tors within extensive computational infrastructures, including
cloud platforms and computer clusters. The versatility and
scalability offered by headless simulation enable the effective
Fig. 7. A series of generated unsafe scenarios [95]. These attackers can help management of large-scale simulation tasks [69].
test the ability of the autonomous vehicle to avoid accidents.
Most existing simulators are built upon game engines or
graphical frameworks [11, 45, 46]. The presence of headless
simulation in open-source simulators is examined in TABLE
behavior patterns [11]. These models are constrained to actions II. Achieving headless simulation necessitates comprehensive
like maintaining a specific distance and stopping, lacking the customizations spanning diverse domains: fine-tuning ren-
flexibility to circumvent obstacles or accelerate through scenes dering protocols, optimizing data transmission and storage
[100, 101, 102]. The notable disparity between the behavioral mechanisms, and implementing performance enhancements
models of simulated traffic participants and their real-world meticulously tailored to the headless simulation paradigm.
counterparts undermines the credibility of simulation experi- This intricate endeavor is pivotal for effectively dissociating
ments in driving decision-making task [32]. the simulator from graphical interfaces, thus enabling seamless
operation within remote server environments.
Indeed, several methods have been devised to enhance the
interaction capabilities of traffic participants [23, 27]. A preva-
E. Multiple-speed Simulation
lent approach involves adding an element of randomness to the
behavior of these participants. The typical procedure entails Multiple-speed simulation refers to accelerating the simu-
the behavior model initially outlining potential choices to lation speed without compromising the accuracy of rendering
navigate the traffic participant to pass the scenario successfully. and physics updates. This feature significantly improves the
Then a trajectory is randomly chosen based on historical data efficiency of simulation testing, enabling engineering tests that
distributions accumulated from analogous situations. These typically take several months in the real world to be completed
data distributions can be represented through explicit formulas in half or even less time. The implementation of multiple-
or harnessed through neural networks [3, 103, 104]. This tech- speed simulation follows a similar framework to high frame-
nique aids in imbuing traffic participants with unpredictability, rate simulation. However, based on investigations of open-
contributing to a more authentic and varied simulation environ- source and actively maintained simulators, it is apparent that
ment. The flexibility and complexity of generating interactive the frame rate rarely exceeds 60 Hz (as shown in TABLE II),
traffic participants still have much room for improvement. and most simulators do not support multiple-speed simulation.
The frame rate further drops when the simulation grows
Besides improving traffic participant controlling algorithms, complex, such as high traffic density and adverse weather,
integrating algorithms with the simulators is another pivotal which indicates a pressing need for performance optimization.
challenge. Interactive driving strategies tend to exhibit more Here are some feasible directions for improvement:
aggressiveness, calling for an efficient way to detect traffic rule 1) Algorithm optimization: Improve the data structures and
violations and collisions. One possible solution is to adopt a algorithms for collision detection, physics update, rendering,
multi-agent driving strategy model to substitute the traditional and other computational tasks to increase efficiency.
traffic participants. In this way, all vehicles within the traffic 2) Code optimization: Optimize the code of the simulator
scenario assume the role of controllable entities governed to eliminate unnecessary computations, memory usage, and
by the driving strategy model. This approach enhances the delays, thereby improving overall performance.
intelligence of traffic participants while concurrently ensuring
3) Data compression: Use efficient data compression al-
that the driving strategy models are meticulously developed
gorithms to reduce storage and transmission overhead of
and rigorously tested, so it has gained support from some
simulation data, improving simulation efficiency.
of the simulators and algorithms [11, 46, 57, 105]. The
4) Level of details: For certain scenarios or testing ob-
other solution to accelerate simulation and save computation
jectives, the level of detail can be selectively expressed to
resources will be discussed in Section VI-E.
accelerate the simulation speed. This can be achieved through
Moreover, open-source simulators should offer an interface simplifying models, lowering texture resolution, or reducing
for users to customize traffic participants. By equipping simu- the precision of physics calculations, etc [71].
lators with a dependable traffic event detector and a framework 5) Parallel and distributed computation: To accelerate sim-
to tailor behavior models, the creativity of the open-source ulation speed, a viable solution is to leverage the power of
community can facilitate realizing more interactive traffic multi-core processors and distributed computation by distribut-
participant controllers through collaborative contributions. ing tasks across multiple processing units [72].
11

TABLE II
T HE HARDWARE REQUIREMENTS AND PERFORMANCE OF THE FREE AND OPEN - SOURCE SIMULATORS THAT ARE ACTIVELY MAINTAINED

Minimal hardware requirements Headless Parallel Distributed Maximal


Simulator
Disk space Memory GPU simulation computation computation FPS
SUMO 1 GB 512 MB None 120 Hz
√ √
Webots 1 GB 2 GB 512 MB √ √ √ 60 Hz
Gazebo 500 MB 4 GB 1 GB 500 Hz

CarRacing 1 MB 128 MB None 60 Hz
VDrift 512 MB 1GB None √ √ 60 Hz
Apollo 2 GB 16 GB None 60 Hz
CommonRoad 32 MB 2 GB None 60 Hz
highway-env 1 MB 128 MB None 60 Hz
SMARTS 128 MB 2 GB None 60 Hz
√ √ √
CARLA 20 GB 8 GB 8GB 60 Hz

VII. E VALUATION M ETRICS This method does not directly evaluate data generation perfor-
After comprehending the critical issues of simulators and mance but involves using intermediate algorithms to compare
their potential remedies, it is essential to employ qualitative the synthesized data distributions. The underlying theory is
and quantitative metrics for evaluating simulator performance as follows: assume that real-world data follows a distribution
in specific aspects. This section reviews and explains the pattern D, data generated by method I follows distribution D′ ,
evaluation metrics for assessing simulation fidelity, organized and data generated by method II follows distribution D′′ . By
by the metrics’ mechanisms. Given the limited number of training intermediate algorithms α and β on datasets produced
research addressing critical issues and the divergence in the by method I and method II, respectively, the prediction
metric selection, this section will only explain the indicators accuracy can be assessed. Upon transferring the models to real-
employed across multiple research papers in detail. The evalu- world datasets, if the accuracy of method I surpasses that of
ation of simulation efficiency, being more straightforward and method II, it can be inferred that the synthesized distribution
unified, will be condensed and presented in the last subsection. D′ is closer to the real-world data distribution D, indicating
the superiority of method I over method II.
A. Visualization Numerous algorithms for synthesizing images [64, 84], lidar
points [65], radar [81] and other sensor data have utilized
Visualization is a straightforward and intuitive approach
SOTA perception algorithms as intermediaries to demonstrate
to assess the realism of image synthesis. Early research on
their effectiveness. The widely used accuracy metrics for
adverse weather simulation commonly involved comparing
intermediate perception algorithms include Average Precision
synthesized images with real-world images and outputs of
[63, 83], mean Average Precision [64], Intersection of Union
other algorithms (as shown in Fig. 5) [64, 79]. Visualization
(IoU) [80, 97], and F1 score. These metrics connect synthe-
is equally applicable to synthesized lidar points and radar
sizing methods’ performance and practical tasks, making them
data (as shown in Fig. 4) [65, 81]. While visualization alone
valuable for real-world applications. However, this evaluation
may not be sufficient to fully validate the significance of data
method also has certain drawbacks. The main limitation is
synthesis algorithms, it remains a crucial aspect of evaluation.
the time cost, as the evaluation process requires significant
Quantitative metrics can test the performance of data synthesis
data processing and model training [76]. Another concern is
algorithms in specific aspects, but they may not capture the
the lack of theoretical research to justify the statistical signif-
finer details and realism of image synthesis. By visually
icance. It remains unclear how much percentile of accuracy
presenting the differences between synthesized images and
the media algorithm should reach to prove one synthesizing
real-world images, visualization aids researchers in better
method is an advantage over others convincingly.
understanding the output and identifying potential issues for
further improvements.
Despite its advantages, visualization also has limitations.
First, it is a subjective method, so every individual may have C. Accuracy-based Metric
a unique perception [85]. Second, visual results lack specific Using accuracy metrics for map reconstruction when ground
quantitative metrics to directly measure the difference between truth is available is a viable approach. When the reconstructed
synthesized and real-world images. Therefore, quantitative map contains models or structured organization, metrics like
metrics and visual results are both necessary for a compre- IoU or F1 scores can be used to compare the accuracy of
hensive assessment of data synthesis algorithms. the reconstruction with real-world data [68, 106]. For map
reconstruction models based on NeRF and other neural kernel-
B. Distribution-based Metric based methods, image quality assessment metrics such as Peak
Assessing synthesizing algorithms through distribution- Signal-to-Noise Ratio (PSNR) and Structural Similarity Index
based metrics is a practical and widely applicable approach. (SSIM) can also be introduced [67, 93].
12

PSNR is calculated using the Mean Squared Error (MSE) VIII. C ONCLUSIONS
formula and is generally expressed as
This paper presents a comprehensive overview of the history
P SN R = 20 · log10 (M AXI ) − 10 · log10 (M SE) (1) of simulators for autonomous driving over the past three
decades. The timeline is divided into three distinct phases:
where M AXI represents the maximum possible pixel value in the specialized development period (1990s-2002), the gap pe-
the image, and M SE is the difference between the generated riod (2002-2013), and the comprehensive development period
image and the ground truth. A higher PSNR value indicates (2013-present). A tendency toward comprehensive functional-
smaller noise and better reconstruction results. ity and open-source accessibility is detected in the evolution
SSIM, on the other hand, assesses the algorithm’s quality process. After investigating influential commercial and open-
by luminance, contrast, and structure and is represented as source simulators, they are classified into five specialized
classes: traffic flow simulators, vehicle dynamics simulators,
SSIM (x, y) = [l(x, y)α · c(x, y)β · s(x, y)γ ] (2)
scene editors, sensor data synthesizers, and driving strategy
where x and y are the generated image and the real-world validators. Additionally, simulators that enclose multiple func-
image, and l(x, y), c(x, y), and s(x, y) are the evaluation tionalities are grouped as comprehensive simulators.
functions for brightness, contrast, and structure, respectively. After obtaining a systematic understanding of the simula-
α, β, γ are adjustable parameters [107]. tors, this paper identifies the critical issues that defect the
Despite their advantages, accuracy-based evaluation meth- usage of simulators by observing the open-source simula-
ods have some drawbacks. They often require manual an- tors. The primary concerns are two aspects, data fidelity,
notations on real-world data, which can be time-consuming and simulation efficiency. This paper argues that addressing
and labor-intensive [106]. Additionally, accurately matching challenges related to adverse weather simulation, automatic
corresponding points between synthetic and real-world data map reconstruction, and interactive traffic participants will
is crucial to avoid inaccuracies in the evaluation results due greatly enhance the realism of simulators. Additionally, devel-
to matching errors [88]. Furthermore, accuracy-based metrics oping a headless simulation pipeline and achieving high frame
only provide specific insights into data quality and may not rates can significantly improve the efficiency of simulators.
comprehensively evaluate algorithm performance. A combina- This paper provides a clear definition of the critical issues,
tion of different evaluation methods is often necessary for a emphasizes their importance, and outlines feasible solutions as
comprehensive assessment. valuable guidelines for researchers who tend to use or improve
simulators. Moreover, the paper explores both qualitative and
quantitative metrics to evaluate the effectiveness of resolving
D. Metrics for Simulation Efficiency these challenges within simulators, elucidating the principles
Validating headless simulation is a pure engineering task. and scope of applicability of these metrics.
It involves running the simulator across various operating This paper is written as a call for more attention to the
systems and environmental configurations to confirm its in- simulators for autonomous driving. Continuous and systematic
dependence from the GUI. The assessment process entails improvements in simulators are essential to the success of
numerous iterations of trial and error. The outcome of this autonomous driving systems. As the field progresses, further
testing is a straightforward true-false answer about whether advancements in simulators will fasten the realization and
headless simulation is successfully implemented. application of autonomous driving technology.
The assessment of multiple-speed simulation involves mon-
itoring the simulator’s execution speed and allocated com-
R EFERENCES
putational resources. Tracking execution speed is typically
achieved through plugins and programs, with specifics deter- [1] N. Kalra and S. M. Paddock, “Driving to safety: How
mined by the simulator’s programming language and frame- many miles of driving would it take to demonstrate au-
work architecture. In terms of computational resource usage, tonomous vehicle reliability?” Transportation Research
critical metrics to monitor are CPU utilization, memory al- Part A: Policy and Practice, vol. 94, pp. 182–193, 2016.
location, GPU utilization, and GPU memory allocation. For [2] A. Shafaei, J. J. Little, and M. Schmidt, “Play and learn:
high-performance simulators supporting parallel processing, Using video games to train computer vision models,”
performance evaluation often employs metrics like the speedup arXiv preprint arXiv:1608.01745, 2016.
ratio S and speedup efficiency E, defined as follows: [3] Q. Zhang, Y. Gao, Y. Zhang, Y. Guo, D. Ding, Y. Wang,
P. Sun, and D. Zhao, “Trajgen: Generating realistic
serial execution time and diverse trajectories with reactive and feasible agent
S= (3)
parallel execution time behaviors for autonomous driving,” IEEE Transactions
on Intelligent Transportation Systems, vol. 23, no. 12,
S pp. 24 474–24 487, 2022.
E= (4)
P [4] S. Wang, R. Stern, and M. W. Levin, “Optimal con-
where P denotes the number of processors. The speedup trol of autonomous vehicles for traffic smoothing,”
efficiency E is expected to be greater or equal to 1. If E ≪ IEEE Transactions on Intelligent Transportation Sys-
1, the optimization is invalid [108]. tems, vol. 23, no. 4, pp. 3842–3852, 2021.
13

[5] Y. Wang, S. Zhao, R. Zhang, X. Cheng, and L. Yang, har, K. Rosaen, and R. Vasudevan, “Driving in the
“Multi-vehicle collaborative learning for trajectory pre- matrix: Can virtual worlds replace human-generated
diction with spatio-temporal tensor fusion,” IEEE annotations for real world tasks?” arXiv preprint
Transactions on Intelligent Transportation Systems, arXiv:1610.01983, 2016.
vol. 23, no. 1, pp. 236–248, 2020. [21] S. Grigorescu, B. Trasnea, T. Cocias, and G. Macesanu,
[6] IEEE Transactions on Intelligent Transportation Sys- “A survey of deep learning techniques for autonomous
tems, “Ieee transactions on intelligent transportation driving,” Journal of Field Robotics, vol. 37, no. 3, pp.
systems all issues 2022,” https://ieeexplore.ieee.org/xpl/ 362–386, 2020.
issues?punumber=6979&isnumber=10175858, accessed [22] C. Katrakazas, M. Quddus, W.-H. Chen, and L. Deka,
Aug. 14, 2023. “Real-time motion planning methods for autonomous
[7] The Waymo Team, “Simulation city: Introducing on-road driving: State-of-the-art and future research
waymo’s most advanced simulation system yet for directions,” Transportation Research Part C: Emerging
autonomous driving,” https://waymo.com/blog/2021/06/ Technologies, vol. 60, pp. 416–442, 2015.
SimulationCity.html, accessed Aug. 14, 2023. [23] W. Schwarting, J. Alonso-Mora, and D. Rus, “Planning
[8] Pony.ai, “Technology,” https://pony.ai/tech?lang=en, ac- and decision-making for autonomous vehicles,” Annual
cessed Aug. 14, 2023. Review of Control, Robotics, and Autonomous Systems,
[9] DiDi, “Didi’s autonomous driving,” https: vol. 1, pp. 187–210, 2018.
//www.didiglobal.com/science/intelligent-driving, [24] L. Claussmann, M. Revilloud, D. Gruyer, and S. Glaser,
accessed Aug. 14, 2023. “A review of motion planning for highway autonomous
[10] A. Sinha, S. Chand, V. Vu, H. Chen, and V. Dixit, driving,” IEEE Transactions on Intelligent Transporta-
“Crash and disengagement data of autonomous vehicles tion Systems, vol. 21, no. 5, pp. 1826–1848, 2019.
on public roads in california,” Scientific data, vol. 8, [25] U. M. Gidado, H. Chiroma, N. Aljojo, S. Abubakar,
no. 1, p. 298, 2021. S. I. Popoola, and M. A. Al-Garadi, “A survey on deep
[11] A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and learning for steering angle prediction in autonomous
V. Koltun, “Carla: An open urban driving simulator,” in vehicles,” IEEE Access, vol. 8, pp. 163 797–163 817,
Conference on robot learning. PMLR, 2017, pp. 1–16. 2020.
[12] Baidu, “Apollo,” https://github.com/ApolloAuto/apollo, [26] A. Tampuu, T. Matiisen, M. Semikin, D. Fishman,
accessed Aug. 14, 2023. and N. Muhammad, “A survey of end-to-end driving:
[13] M. Smith, G. Duncan, and S. Druitt, “Paramics: micro- Architectures and training methods,” IEEE Transactions
scopic traffic simulation for congestion management,” on Neural Networks and Learning Systems, vol. 33,
1995. no. 4, pp. 1364–1384, 2020.
[14] J. Barceló and J. Casas, “Dynamic network simula- [27] L. Wei, Z. Li, J. Gong, C. Gong, and J. Li, “Autonomous
tion with aimsun,” in Simulation approaches in trans- driving strategies at intersections: Scenarios, state-of-
portation analysis: Recent advances and challenges. the-art, and future outlooks,” in 2021 IEEE Interna-
Springer, 2005, pp. 57–98. tional Intelligent Transportation Systems Conference
[15] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz, (ITSC). IEEE, 2021, pp. 44–51.
“Sumo–simulation of urban mobility: an overview,” in [28] J. Zhou, Y. Zhang, S. Guo, and Y. Guo, “A survey
Proceedings of SIMUL 2011, The Third International on autonomous driving system simulators,” in 2022
Conference on Advances in System Simulation. Think- IEEE International Symposium on Software Reliability
Mind, 2011. Engineering Workshops (ISSREW). IEEE, 2022, pp.
[16] D. Feng, A. Harakeh, S. L. Waslander, and K. Di- 301–306.
etmayer, “A review and comparative study on proba- [29] M. Holen, K. Knausgård, and M. Goodwin, “An eval-
bilistic object detection in autonomous driving,” IEEE uation of autonomous car simulators and their appli-
Transactions on Intelligent Transportation Systems, cability for supervised and reinforcement learning,” in
vol. 23, no. 8, pp. 9961–9980, 2021. International Conference on Intelligent Technologies
[17] A. Haydari and Y. Yılmaz, “Deep reinforcement learn- and Applications. Springer, 2021, pp. 367–379.
ing for intelligent transportation systems: A survey,” [30] P. Kaur, S. Taghavi, Z. Tian, and W. Shi, “A survey
IEEE Transactions on Intelligent Transportation Sys- on simulators for testing self-driving cars,” in 2021
tems, vol. 23, no. 1, pp. 11–32, 2020. Fourth International Conference on Connected and Au-
[18] Y. Li, R. Chen, X. Niu, Y. Zhuang, Z. Gao, X. Hu, tonomous Driving (MetroCAD). IEEE, 2021, pp. 62–
and N. El-Sheimy, “Inertial sensing meets machine 70.
learning: Opportunity or challenge?” IEEE Transactions [31] X. Hu, S. Li, T. Huang, B. Tang, and L. Chen, “Sim2real
on Intelligent Transportation Systems, 2021. and digital twins in autonomous driving: A survey,”
[19] E. Yurtsever, J. Lambert, A. Carballo, and K. Takeda, arXiv preprint arXiv:2305.01263, 2023.
“A survey of autonomous driving: Common practices [32] C. Siebke, M. Mai, and G. Prokop, “What do traffic
and emerging technologies,” IEEE Access, vol. 8, pp. simulations have to provide for virtual road safety
58 443–58 469, 2020. assessment? human error modeling in traffic simula-
[20] M. Johnson-Roberson, C. Barto, R. Mehta, S. N. Srid- tions,” IEEE Transactions on Intelligent Transportation
14

Systems, vol. 24, no. 2, pp. 1419–1436, 2022. 2004 IEEE/RSJ International Conference on Intelli-
[33] PTV Group, “Ptv vissim: Multimodel traffic gent Robots and Systems (IROS) (IEEE Cat. No.
simulation software,” https://www.myptv.com/en/ 04CH37566), vol. 3. IEEE, 2004, pp. 2149–2154.
mobility-software/ptv-vissim, accessed Aug. 14, 2023. [49] MathWorks, “Get started with roadrunner,”
[34] K. G. Lim, C. H. Lee, R. K. Y. Chin, K. B. Yeo, https://www.mathworks.com/help/roadrunner/get-
and K. T. K. Teo, “Sumo enhancement for vehicular started-with-roadrunner.html, accessed Aug. 14, 2023.
ad hoc network (vanet) simulation,” in 2017 IEEE [50] Parallel Domain, “Parallel domain,” https:
2nd international conference on automatic control and //paralleldomain.com/, accessed Aug. 14, 2023.
intelligent systems (I2CACIS). IEEE, 2017, pp. 86–91. [51] Nvidia, “Nvidia drive end-to-end platform for
[35] Mechanical Simulation, “Carsim,” https://www.carsim. software-defined vehicles,” https://www.nvidia.com/
com, accessed Aug. 14, 2023. en-us/self-driving-cars/, accessed Aug. 14, 2023.
[36] IPG Automotive, “Solutions for virtual test driving,” [52] A. Amini, T.-H. Wang, I. Gilitschenski, W. Schwarting,
https://ipg-automotive.com/en, accessed Aug. 14, 2023. Z. Liu, S. Han, S. Karaman, and D. Rus, “Vista 2.0:
[37] rFpro, “The world’s most accurate simulation environ- An open, data-driven simulator for multimodal sensing
ment,” https://rfpro.com, accessed Aug. 14, 2023. and policy learning for autonomous vehicles,” in 2022
[38] Siemens, “Simcenter prescan software,” International Conference on Robotics and Automation
https://plm.sw.siemens.com/en-US/simcenter/ (ICRA). IEEE, 2022, pp. 2419–2426.
autonomous-vehicle-solutions/prescan, accessed [53] B. Wymann, E. Espié, C. Guionneau, C. Dimitrakakis,
Aug. 14, 2023. R. Coulom, and A. Sumner, “Torcs, the open racing car
[39] VI-grade, “Driving simulators: An invaluable set simulator,” Software available at http://torcs. source-
of tools to successfully accelerate your develop- forge. net, vol. 4, no. 6, p. 2, 2000.
ment process,” https://www.vi-grade.com/en/solutions/ [54] C. Campbell, “Box2d c++ tutorials - top-down car
driving simulators, accessed Aug. 14, 2023. physics,” http://www.iforce2d.net/b2dtut/top-down-car,
[40] M. Martinez, C. Sitawarin, K. Finch, L. Meincke, accessed Aug. 14, 2023.
A. Yablonski, and A. Kornhauser, “Beyond grand theft [55] F. Kehrle, J. V. Frasch, C. Kirches, and S. Sager,
auto v for training, testing and enhancing deep learning “Optimal control of formula 1 race cars in a vdrift
in self driving cars,” arXiv preprint arXiv:1712.01397, based virtual environment,” IFAC Proceedings Volumes,
2017. vol. 44, no. 1, pp. 11 907–11 912, 2011.
[41] D. Wang, J. Wen, Y. Wang, X. Huang, and F. Pei, “End- [56] Udacity, “Udacity self-driving platform,” https://github.
to-end self-driving using deep neural networks with com/udacity/self-driving-car-sim, accessed: 2023-07-
multi-auxiliary tasks,” Automotive Innovation, vol. 2, 15.
pp. 127–136, 2019. [57] M. Zhou, J. Luo, J. Villella, Y. Yang, D. Rusu, J. Miao,
[42] S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Air- W. Zhang, M. Alban, I. Fadakar, Z. Chen et al.,
sim: High-fidelity visual and physical simulation for “Smarts: Scalable multi-agent reinforcement learning
autonomous vehicles,” in Field and Service Robotics: training school for autonomous driving,” arXiv preprint
Results of the 11th International Conference. Springer, arXiv:2010.09776, 2020.
2018, pp. 621–635. [58] P. Cai, Y. Lee, Y. Luo, and D. Hsu, “Summit: A
[43] G. Rong, B. H. Shin, H. Tabatabaee, Q. Lu, S. Lemke, simulator for urban driving in massive mixed traffic,” in
M. Možeiko, E. Boise, G. Uhm, M. Gerow, S. Mehta 2020 IEEE International Conference on Robotics and
et al., “Lgsvl simulator: A high fidelity simulator Automation (ICRA). IEEE, 2020, pp. 4023–4029.
for autonomous driving,” in 2020 IEEE 23rd Interna- [59] AVSimulation, “Scaner 2023.2 release note,” https://
tional Conference on Intelligent Transportation Systems www.avsimulation.com/scaner-2023-2-release-note, ac-
(ITSC). IEEE, 2020, pp. 1–6. cessed Aug. 14, 2023.
[44] M. Bansal, A. Krizhevsky, and A. Ogale, “Chauffeurnet: [60] Hexagon, “Virtual test drive: Complete tool-chain for
Learning to drive by imitating the best and synthesizing driving simulation applications,” https://hexagon.com/
the worst,” arXiv preprint arXiv:1812.03079, 2018. products/virtual-test-drive, accessed Aug. 14, 2023.
[45] E. Leurent, “An environment for autonomous [61] P. Li, A. Kusari, and D. J. LeBlanc, “A novel traffic sim-
driving decision-making,” https://github.com/eleurent/ ulation framework for testing autonomous vehicles us-
highway-env, accessed Aug. 14, 2023. ing sumo and carla,” arXiv preprint arXiv:2110.07111,
[46] M. Althoff, M. Koschi, and S. Manzinger, “Common- 2021.
road: Composable benchmarks for motion planning on [62] M. Dupuis, M. Strobl, and H. Grezlikowski, “Opendrive
roads,” in 2017 IEEE Intelligent Vehicles Symposium 2010 and beyond–status and future of the de facto
(IV). IEEE, 2017, pp. 719–726. standard for the description of road networks,” in Proc.
[47] O. Michel, “Cyberbotics ltd. webots™: professional of the Driving Simulation Conference Europe, 2010, pp.
mobile robot simulation,” International Journal of Ad- 231–242.
vanced Robotic Systems, vol. 1, no. 1, p. 5, 2004. [63] A. Von Bernuth, G. Volk, and O. Bringmann, “Simulat-
[48] N. Koenig and A. Howard, “Design and use paradigms ing photo-realistic snow and fog on existing images for
for gazebo, an open-source multi-robot simulator,” in enhanced cnn training and evaluation,” in 2019 IEEE
15

Intelligent Transportation Systems Conference (ITSC). [77] D. Dunkerley, “Rain event properties in nature and in
IEEE, 2019, pp. 41–46. rainfall simulation experiments: a comparative review
[64] M. Tremblay, S. S. Halder, R. De Charette, and J.-F. with recommendations for increasingly systematic study
Lalonde, “Rain rendering for evaluating and improving and reporting,” Hydrological processes: An interna-
robustness to bad weather,” International Journal of tional journal, vol. 22, no. 22, pp. 4415–4435, 2008.
Computer Vision, vol. 129, pp. 341–360, 2021. [78] S. Starik and M. Werman, “Simulation of rain in
[65] M. Hahner, C. Sakaridis, M. Bijelic, F. Heide, F. Yu, videos,” in Texture Workshop, ICCV, vol. 2, 2003, pp.
D. Dai, and L. Van Gool, “Lidar snowfall simulation 406–409.
for robust 3d object detection,” in Proceedings of the [79] C. Wang, Z. Wang, T. Xia, and Q. Peng, “Real-time
IEEE/CVF Conference on Computer Vision and Pattern snowing simulation,” The Visual Computer, vol. 22, pp.
Recognition, 2022, pp. 16 364–16 374. 315–323, 2006.
[66] S. Tan, K. Wong, S. Wang, S. Manivasagam, M. Ren, [80] C. Sakaridis, D. Dai, and L. Van Gool, “Semantic foggy
and R. Urtasun, “Scenegen: Learning to generate real- scene understanding with synthetic data,” International
istic traffic scenes,” in Proceedings of the IEEE/CVF Journal of Computer Vision, vol. 126, pp. 973–992,
Conference on Computer Vision and Pattern Recogni- 2018.
tion, 2021, pp. 892–901. [81] D.-H. Paek, S.-H. Kong, and K. T. Wijaya, “K-radar: 4d
[67] Z. Li, T. Müller, A. Evans, R. H. Taylor, M. Unberath, radar object detection for autonomous driving in various
M.-Y. Liu, and C.-H. Lin, “Neuralangelo: High-fidelity weather conditions,” Advances in Neural Information
neural surface reconstruction,” in Proceedings of the Processing Systems, vol. 35, pp. 3819–3829, 2022.
IEEE/CVF Conference on Computer Vision and Pattern [82] M. Zhang, Y. Zhang, L. Zhang, C. Liu, and S. Khurshid,
Recognition, 2023, pp. 8456–8465. “Deeproad: Gan-based metamorphic testing and input
[68] Y. Tian, A. Carballo, R. Li, and K. Takeda, “Real-to- validation framework for autonomous driving systems,”
synthetic: Generating simulator friendly traffic scenes in Proceedings of the 33rd ACM/IEEE International
from graph representation,” in 2022 IEEE Intelligent Conference on Automated Software Engineering, 2018,
Vehicles Symposium (IV). IEEE, 2022, pp. 1615–1622. pp. 132–142.
[69] A. Afzal, D. S. Katz, C. L. Goues, and C. S. Timperley, [83] Z. Yang, Y. Chai, D. Anguelov, Y. Zhou, P. Sun,
“A study on the challenges of using robotics simulators D. Erhan, S. Rafferty, and H. Kretzschmar, “Surfelgan:
for testing,” arXiv preprint arXiv:2004.07368, 2020. Synthesizing realistic sensor data for autonomous driv-
[70] S. J. Seyed Aboutorabi and M. H. Rezvani, “An opti- ing,” in Proceedings of the IEEE/CVF Conference on
mized meta-heuristic bees algorithm for players’ frame Computer Vision and Pattern Recognition, 2020, pp.
rate allocation problem in cloud gaming environments,” 11 118–11 127.
The Computer Games Journal, vol. 9, pp. 281–304, [84] V. Mus, at, I. Fursa, P. Newman, F. Cuzzolin, and
2020. A. Bradley, “Multi-weather city: Adverse weather stack-
[71] B. Varga, D. Doba, and T. Tettamanti, “Optimizing vehi- ing for autonomous driving,” in Proceedings of the
cle dynamics co-simulation performance by introducing IEEE/CVF International Conference on Computer Vi-
mesoscopic traffic simulation,” Simulation Modelling sion, 2021, pp. 2906–2915.
Practice and Theory, vol. 125, p. 102739, 2023. [85] Y. Li, Z.-H. Lin, D. Forsyth, J.-B. Huang, and S. Wang,
[72] A. Rousset, B. Herrmann, C. Lang, and L. Philippe, “A “Climatenerf: Physically-based neural rendering for ex-
survey on parallel and distributed multi-agent systems treme climate synthesis,” arXiv e-prints, pp. arXiv–
for high performance computing simulations,” Com- 2211, 2022.
puter Science Review, vol. 22, pp. 27–46, 2016. [86] C. Früh and A. Zakhor, “An automated method for
[73] J. Vargas, S. Alsweiss, O. Toker, R. Razdan, and large-scale, ground-based city model acquisition,” Inter-
J. Santos, “An overview of autonomous vehicles sensors national Journal of Computer Vision, vol. 60, pp. 5–24,
and their vulnerability to weather conditions,” Sensors, 2004.
vol. 21, no. 16, p. 5397, 2021. [87] N. Snavely, S. M. Seitz, and R. Szeliski, “Photo tourism:
[74] C. Goodin, D. Carruth, M. Doude, and C. Hudson, exploring photo collections in 3d,” in ACM siggraph
“Predicting the influence of rain on lidar in adas,” 2006 papers, 2006, pp. 835–846.
Electronics, vol. 8, no. 1, p. 89, 2019. [88] S. Agarwal, Y. Furukawa, N. Snavely, I. Simon, B. Cur-
[75] S. Zang, M. Ding, D. Smith, P. Tyler, T. Rakotoarivelo, less, S. M. Seitz, and R. Szeliski, “Building rome in a
and M. A. Kaafar, “The impact of adverse weather day,” Communications of the ACM, vol. 54, no. 10, pp.
conditions on autonomous vehicles: How rain, snow, 105–112, 2011.
fog, and hail affect the performance of a self-driving [89] F. Williams, M. Trager, J. Bruna, and D. Zorin, “Neural
car,” IEEE Vehicular Technology Magazine, vol. 14, splines: Fitting 3d surfaces with infinitely-wide neural
no. 2, pp. 103–111, 2019. networks,” in Proceedings of the IEEE/CVF Conference
[76] W. Li, Q. Zhang, J. Zhang, Z. Huang, X. Tian, on Computer Vision and Pattern Recognition, 2021, pp.
and D. Tao, “Toward real-world single image derain- 9949–9958.
ing: A new benchmark and beyond,” arXiv preprint [90] J. Pan, X. Han, W. Chen, J. Tang, and K. Jia, “Deep
arXiv:2206.05514, 2022. mesh reconstruction from single rgb images via topol-
16

ogy modification networks,” in Proceedings of the [103] A. Kesting, M. Treiber, and D. Helbing, “General
IEEE/CVF International Conference on Computer Vi- lane-changing model mobil for car-following models,”
sion, 2019, pp. 9964–9973. Transportation Research Record, vol. 1999, no. 1, pp.
[91] A. Boulch and R. Marlet, “Poco: Point convolution 86–94, 2007.
for surface reconstruction,” in Proceedings of the [104] F.-Y. Wang, “Parallel control and management for intel-
IEEE/CVF Conference on Computer Vision and Pattern ligent transportation systems: Concepts, architectures,
Recognition, 2022, pp. 6302–6314. and applications,” IEEE Transactions on Intelligent
[92] J. Huang, Z. Gojcic, M. Atzmon, O. Litany, S. Fidler, Transportation Systems, vol. 11, no. 3, pp. 630–638,
and F. Williams, “Neural kernel surface reconstruction,” 2010.
in Proceedings of the IEEE/CVF Conference on Com- [105] S. Shalev-Shwartz, S. Shammah, and A. Shashua, “Safe,
puter Vision and Pattern Recognition, 2023, pp. 4369– multi-agent, reinforcement learning for autonomous
4379. driving,” arXiv preprint arXiv:1610.03295, 2016.
[93] M. Tancik, V. Casser, X. Yan, S. Pradhan, B. Milden- [106] A. Knapitsch, J. Park, Q.-Y. Zhou, and V. Koltun,
hall, P. P. Srinivasan, J. T. Barron, and H. Kretzschmar, “Tanks and temples: Benchmarking large-scale scene
“Block-nerf: Scalable large scene neural view synthe- reconstruction,” ACM Transactions on Graphics (ToG),
sis,” in Proceedings of the IEEE/CVF Conference on vol. 36, no. 4, pp. 1–13, 2017.
Computer Vision and Pattern Recognition, 2022, pp. [107] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simon-
8248–8258. celli, “Image quality assessment: from error visibility
[94] A. Mondal, P. Tigas, and Y. Gal, “Real2sim: Automatic to structural similarity,” IEEE Transactions on Image
generation of open street map towns for autonomous Processing, vol. 13, no. 4, pp. 600–612, 2004.
driving benchmarks,” in Machine Learning for Au- [108] S. Sahni and V. Thanvantri, “Parallel computing: Per-
tonomous Driving Workshop at the 34th Conference formance metrics and models,” IEEE Parallel and Dis-
on Neural Information Processing Systems (NeurIPS), tributed Technology, vol. 4, no. 1, pp. 43–56, 1996.
2020.
[95] Z. Ghodsi, S. K. S. Hari, I. Frosio, T. Tsai, A. Troccoli,
S. W. Keckler, S. Garg, and A. Anandkumar, “Gener-
ating and characterizing scenarios for safety testing of Yueyuan LI received a Bachelor’s degree in Electri-
autonomous vehicles,” in 2021 IEEE Intelligent Vehicles cal and Computer Engineering from the University
of Michigan-Shanghai Jiao Tong University Joint
Symposium (IV). IEEE, 2021, pp. 157–164. Insitute, Shanghai, China, in 2020. She is working
[96] J. Gao, C. Sun, H. Zhao, Y. Shen, D. Anguelov, C. Li, towards a Ph.D. degree in Control Science and
and C. Schmid, “Vectornet: Encoding hd maps and Engineering from Shanghai Jiao Tong University.
Her main fields of interest are the security of the
agent dynamics from vectorized representation,” in Pro- autonomous driving system and driving decision-
ceedings of the IEEE/CVF Conference on Computer Vi- making. Her current research activities include driv-
sion and Pattern Recognition, 2020, pp. 11 525–11 533. ing decision-making models, driving simulation, and
virtual-to-real model transferring.
[97] A. Kar, A. Prakash, M.-Y. Liu, E. Cameracci, J. Yuan,
M. Rusiniak, D. Acuna, A. Torralba, and S. Fidler,
“Meta-sim: Learning to generate synthetic datasets,” in
Proceedings of the IEEE/CVF International Conference
Wei YUAN received his Master’s and Ph.D. degrees
on Computer Vision, 2019, pp. 4551–4560. in Automation from Shanghai Jiao Tong University,
[98] M. Haklay and P. Weber, “Openstreetmap: User- Shanghai, China, in 2017 and 2021, respectively.
generated street maps,” IEEE Pervasive computing, Presently, he is a postdoctoral researcher at Shanghai
Jiao Tong University.
vol. 7, no. 4, pp. 12–18, 2008. His main fields of interest are autonomous driving
[99] F. Poggenhans, J.-H. Pauls, J. Janosovits, S. Orf, system, computer vision, deep learning, and vehicle
M. Naumann, F. Kuhnt, and M. Mayr, “Lanelet2: A control. His current research activities include end-
to-end learning-based vehicle control and decision-
high-definition map framework for the future of auto- making.
mated driving,” in 2018 21st International Conference
on Intelligent Transportation Systems (ITSC). IEEE,
2018, pp. 1672–1679.
[100] S. Krauß, “Microscopic modeling of traffic flow: Inves-
tigation of collision free vehicle dynamics,” 1998. Weihao YAN received a Bachelor’s Adegree in
[101] J. Erdmann, “Sumo’s lane-changing model,” in Model- Automation from Shanghai Jiao Tong University,
ing Mobility with Open Data: 2nd SUMO Conference Shanghai, China, in 2020. He is working towards
a Ph.D. degree in Control Science and Engineering
2014 Berlin, Germany, May 15-16, 2014. Springer, from Shanghai Jiao Tong University.
2015, pp. 105–123. His main fields of interest are autonomous driving
[102] S. Panwai and H. Dia, “Comparative evaluation of mi- system, computer vision, and domain adaptation.
His current research activities include virtual-to-real
croscopic car-following behavior,” IEEE Transactions transfer learning, scene segmentation, and founda-
on Intelligent Transportation Systems, vol. 6, no. 3, pp. tion models.
314–325, 2005.
17

Qiyuan SHEN received a Bachelor’s degree in Au-


tomation from Northeastern University, Shenyang,
China, in 2022. He is currently working towards a
Ph.D. degree in Control Science and Engineering at
Shanghai Jiao Tong University.
His main fields of interest are the localization of
autonomous driving systems and SLAM. His current
research activities include multi-modal mapping,
cross-modal localization, and geometrical calibration
for vision and LiDAR.

Chunxiang WANG received a Ph.D. degree in


Mechanical Engineering from Harbin Institute of
Technology, China, in 1999. She is currently an
associate professor in the Department of Automation
at Shanghai Jiao Tong University, Shanghai, China.
She has been working in the field of intelligent
vehicles for more than ten years and has participated
in several related research projects, such as European
CyberC3 project, ITER transfer cask project, etc.
Her research interests include autonomous driving,
assistant driving, and mobile robots.

Ming YANG received his Master’s and Ph.D. de-


grees from Tsinghua University, Beijing, China, in
1999 and 2003, respectively. Presently, he holds
the position of Distinguished Professor at Shanghai
Jiao Tong University, also serving as the Director
of the Innovation Center of Intelligent Connected
Vehicles. Dr. Yang has been engaged in the research
of intelligent vehicles for more than 25 years.

You might also like