Teng Liu,,Hong Wang,Bin Tian, Yunfeng Ai,and Long Chen

Abstract—In this paper,a new paradigm named parallel distance is presented to measure the data information in parallel driving system. As an example, the core variables in the parallel driving system are measured and evaluated in the parallel distance framework. First, the parallel driving 3.0 system included control and management platform, intelligent vehicle platform and remote-control platform is introduced.Then,Markov chain(MC) is utilized to model the transition probability matrix of control commands in these systems. Furthermore, to distinguish the control variables in artificial and physical driving conditions,different distance calculation methods are enumerated to specify the differences between the virtual and real signals. By doing this, the real system can be guided and the virtual system can be im-proved. Finally,simulation results exhibit the merits and multiple applications of the proposed parallel distance framework.

I.Introduction

CONNECTED automated technologies experience a rapid development due to the potential to make more safer and intelligent vehicles and make more reliable and efficient transportation[1],[2].More attention has been paid to the connected automated vehicles(CAVs)and studies on decision-making and planning,perception and location,communication and control are conducted in recent years[3],[4].How to achieve full automation in the urban and high-way roads is still an open question that need to be addressed by academic researchers,industrial manufacturers and governments[5]–[7].

Parallel driving seems a promising solution to realize highly automated driving safely and efficiently.The ACP theory was formulated and presented by Fei-Yue Wang since2004,which indicated modeling by artificial societies(A),analyzing by computational experiments(C)and controlling via parallel execution(P)[8]–[11].Based on this concept,parallel driving is presented as a paradigm of connected automated driving with three parallel worlds.The first level is the physical world,the second level is the mental world and the third level is the artificial world,see Fig.1 as an illustration [12].

In the physical world, the vehicle is handled by the human driver in the real driving condition.Different automation levels can be adopted in this world depend on vehicle technologies and driver preference[12].In the mental world,the driver cognition and behavior are recognized and analyzed,such as attention and neuromuscular attributes[13].Furthermore,two layers co-exist in the artificial world,wherein the first layer is software-defined artificial vehicle and artificial driver,and the second is the information layer consists of sensor,location, people,environment and so on[14].

The vehicle and driver are modeled via artificial societies(A) to imitate the physical driving[15],[16].Computational experiments(C)are constructed in the physical and artificial worlds,respectively[17],[18].Through parallel execution in these three worlds,the artificial world can be improved and the physical world can be guided[19]–[21].By doing this,many controls and signals in different worlds are calculated in parallel.Thus,the differences between these variables need to be measured in order to instruct the practical applications[22],[23].However,an integrated and particular measurement approach has not been involved in the parallel driving framework.

Based on this requirement, this work proposes a new paradigm named parallel distance to quantify the differences between the physical and artificial variables in the parallel driving system.Firstly,the novel parallel driving 3.0 system is described,which incorporates the control and management platform,intelligent vehicle platform and remote-control platform.Secondly,to represent the statistical characteristics of the control commands in these systems,Markov chain(MC)is used to model the transition probability matrix (TPM)of these variables.Thirdly,the parallel distance framework is introduced to measure the real and virtual signals.Multiple distance computation methods are compared to identify the calculated performance and speed.Finally,simulation is designed to show the merits and multiple applications of the proposed parallel distance framework.

Fig.1.Parallel driving configuration.

Three potential perspectives are contributed in this paper:1)a parallel distance framework is presented to measure the differences between the real world and artificial world;2)multiple distance calculation techniques are employed to quantify the virtual and real signals;3)the parallel driving system is introduced in detail for the first time.The main difference between the parallel driving and connected automated driving is that the built artificial model could guide and manage the real autonomous vehicles.The parallel distance presented in this paper is one choice to realize the evaluation and comparison process.By doing this, the real driving vehicles can be guided through supervised control and the virtual modeling can be improved via feedback control.For all we know,this article is the pioneer to discuss measurement method in the parallel driving system.

The following paper is arranged as follow:the parallel driving 3.0 system is depicted in Section II.Section III indicates the parameters of automated vehicle and MC model.Section IV describes the parallel distance framework and multiple distance computation methods.In Section V, the virtual and real control actions are depicted and simulations are designed to validate the presented method,and results are analyzed.Finally,Section VI describes the conclusions.

II.Parallel Driving 3.0 System

The parallel driving 3.0 system includes the control and management platform,intelligent vehicle platform and remote-control platform,as shown in Fig.2.The control and management platform is used to monitor and decide the control switching,the effect of the remote-control platform is managing the intelligent vehicle via driving simulator and the intelligent vehicle can work in different automation levels.As the intelligent vehicle cannot control itself normally, the remote-control platform will take over and act as the telecontrol system.Also, the control switching will happen as the control and management platform recognizes the state of intelligent vehicle is unusual.

Fig.2.Architecture of the parallel driving 3.0 system.

A. Intelligent Vehicle Platform

There are three intelligent vehicles in the parallel driving 3.0 system.They are all equipped with two cameras,laser radar,human machine interface(HM I),emergency stop(E-stop),industrial personal computer(IPC),wireless communication and additional equipment.The sketch of the vehicle is depicted in Fig.3.

The perception,decision-making,planning and control modules are integrated in the IPC.Based on the environment and cartographic information,the intelligent vehicle can manage its motion by itself.Also,the vehicle can receive the control command from the remote-control platform.Thus,there is a control switching module on board,which can communicate with the control and management platform to decide the appropriate control command.

B. Remote-Control Platform

Fig.3.Construction of the intelligent vehicle platform.

The remote-control platform consists of two parts,the driving simulator and the human operator.The driving simulator is described in Fig.4,which contains the viewing screen,steering wheel,accelerator pedal, braking pedal and gear lever.The driving conditions of the virtual vehicle in the artificial world and the virtual mapping modeling are shown in the viewing screen. Also,the camera videos of the real vehicle in the physical world,the real environment and cartographic information and the HIM information are all transferred through 4G network and displayed on the viewing screen.

Fig.4.Sketch of the remote-control platform.

By monitoring the driving conditions of multiple intelligent vehicles on the viewing screen, the human operator can conduct the special vehicle to realize traffic diversion.The human operator will take over the intelligent vehicle in two cases.First,the human operator discerns the abnormal state of vehicle from the viewing screen and takes over the vehicle proactively.Second, the control and management platform would send the takeover command to the human operator.By doing this,several intelligent vehicles will run safely and efficiently.

C.Control and Management Platform

The main components of the control and management platform are artificial simulation equipment,IPC,touch viewing screen, notebook computer,interchanger,cloudy server and image splicer.The function of these components is depicted in Fig.5.

Fig.5.Schematic diagram of the control and management platform.

These are two parts co-exist in the control and management platform.The first one is the software-defined artificial system,which includes the modeling of the virtual driver and virtual vehicle, the modeling of the sensors and the modeling of the traffic environment.The second one is the computational experiment system, wherein the big data,deep learning,parallel reinforcement learning[24]and network communication technologies are applied to handle the data information.

Besides recording and reserving the real driving conditions of the physical vehicle,the control and management platform can also operate the virtual vehicle in various artificial scenarios.Then,the artificial system can train and learn more knowledge than the physical system,and this knowledge will be used to predict the conditions in the real environment and guide the real intelligent vehicle.Finally,the physical system can be guided and the artificial system can be improved.

III.Markov Chain Based Upda ting Algorithm

The main parameters of the automated vehicle and the Markov chain(MC)-based online updating algorithm for transition probability matrix(TPM)are introduced in this section.First,the configuration parameters of the automated vehicles in intelligent vehicle platform are depicted in detail.Also,the data collection equipment is given.Furthermore,to utilize the control commands of the real and virtual systems in real-time,an online updating algorithm is proposed to update the TPM of different control actions.

A.Vehicle Parameters and Data Collection

TABLE I Main Configuration Parameters of Automated Vehicle

IV.Parallel Distance Framework

An artificial virtual intelligent vehicle is established in the parallel driving 3.0 system to imitate the autonomous run of the physical vehicle.The control commands come from the artificial world,the driving simulator and the onboard IPC can guide the physical intelligent vehicle,respectively.Thus,the differences between these commands need to be quantified.The parallel distance framework is introduced in this section to measure the relevant differences.

wherenis the total number of the special signal sequence,πij(P1)denotes the elementπijin theP1.Φis a mapping function to represent the element in TPM by a relevant variable in RKHS.Hence,MMD means the distance for the average value of the elements in TPM in RKHS.

C. Earth Mover’s Distance

In mathematics, the earth mover’s distance(EMD)is defined as a measure of the distance between two probability distributions[32].It is interpreted as the Wasserstein metric.ForP1andP2, the EMD is calculated as follows:

wheredijis the distance of the corresponding elements inP1andP2,fijis the divisor between the norm of two elements.They are given by the transition probabilityπijas

These distance measures are applied to quantify the differences of control commands in the artificial and physical worlds of parallel driving 3.0 system.The comparison process is implemented in MATLAB using the toolbox described in[33].Taking the accelerator pedal and steering wheel angle signals as an example,the next section discusses the merits and the multiple applications of the parallel distance framework.

V.Results and Discussion

For automated driving,the accelerator pedal and steering wheel angle signals are the most important and direct commands to operate the intelligent vehicle[34],[35].In the parallel driving 3.0 system,the control actions from the artificial world can be used to guide the vehicle in the physical world,especially in the situations that would not happen frequently.Furthermore, the control actions from human driver via the driving simulator can improve the modeling accuracy of the artificial world.Thus,the differences between the virtual and real signals need to be computed for multiple applications.

A. Normal Operation of Parallel Driving

The normal flow chart of the parallel driving 3.0 system is shown in Fig.7.The intelligent vehicles are communicated with the control and management platform via the cloud platform,the control and management platform is contacted with the driving simulator by IPC,and the cloud platform can control the driving simulator directly.

As the physical world and artificial world of the parallel3.0 system run normally,the control actions in these two worlds are nearly the same.Fig.8 depicts the TPMs of the accelerator pedal signal in different worlds.It is obvious that these TPMs are similar,which indicates these signals contain analogical statistical characteristics.Furthermore,Fig.9 shows the online updating of the TPM for the accelerator pedal signal.Based on the MC model,the TPM can be changed in real-time according to the variation of this signal.This figure shows the TPMs related to the current and past 1000 s.The differences indicates that the control signals varied acutely in this time interval.

Different distance computation methods in Section IV are applied to compare these control signals,as shown in Fig.10.The distance comparisons for different methods are all small and do not exceed the threshold value,which decides the takeover of parallel driving system.Also,these small distance comparisons mean the operation of the physical and artificial vehicles is similar in the same driving situation.The next subsections discuss the other applications of the parallel distance framework.

B.Guiding Physical Intelligent Vehicle

In some non-normal situations, the on-board controller cannot handle the appropriate control signals,for example the emergency and hostile environment.However,the artificial world can simulate these situations many times to derive the correct signals.Then,the intelligent vehicle can be guided by downloading and applying these virtual control signals from the control and management platform.

The accelerator pedal signal is selected as an example to evaluate that the virtual signal can guide the real vehicle and it is dispersed as[0:0.125:1].The virtual and real accelerator pedal signals in the parallel driving 3.0 system are shown in Fig.11.The virtual signal comes from the control and management platform and the real signal is collected from the intelligent platform.The TPMs corresponding to these signals are calculated by (1)in Section III-A.Three distance computation approaches of the probability distribution in Section III are applied to compare the differences between the virtual and real signals.

Fig.12 depicts the various distance distributions,wherein the TPMs are formulated and compared every 100 s.By setting a threshold value a prior,the time interval when the real signal should be replaced by the virtual one can be decided.It is obvious that different distance calculation method results in different time interval.The time intervals in JSD are[300,500]and[600,700],in MMD are[400,600]and in EMD are[300,700].Hence,the suitable method needs to be chosen according to special driving situations and computation speed in practical applications.

C. Improving Artificial World Modeling

Fig.7.Normal interaction flow chart of the parallel driving 3.0 system.

Fig.8.TPM of real and virtual signals for accelerator pedal.

Fig.9.Updating process of TPM at different time instant.

The real driving traffic environments are complicated and changeable,for example the pedestrian, bicycle and road works could influence the automated driving extremely.Sometimes,the on-board and artificial controllers cannot handle these unexpected factors.Therefore,the human operator would take over the physical intelligent vehicle via the driving simulator.Then, the learning-based methods in the IPC can learn this human experience to improve the artificial world modeling gradually.

Fig.10.Distance comparison for accelerator pedal signal in normal operation.

Fig.11.Real and virtual accelerator pedal signal for comparison.

Choosing the steering wheel angle signal as an example,the real action from the driving simulator and the virtual action from the artificial world are sketched in Fig.13.It is normalized and dispersed as[?1:0.25:1].At time interval [100,200],the control actions are different due to the variational traffic environment.The artificial world need to learn the human experience to update the modeling.To quantify these differences,the distance comparisons in the parallel distance framework are described in Fig.14.

Fig.12.Three distance comparison methods for accelerator pedal signal.

Fig.13.Human and virtual steering wheel angle signal for comparison.

Fig.14.Three distance comparison methods for steering wheel angle signal.

Also,the TPMsP1andP2are compared every 100 s.The threshold value means that the artificial action is not suitable when the distance surpasses the limit.Then, the control actions from the driving simulator should operate the physical intelligent vehicle.The artificial modeling could be improved by learning from this experience,and it can handle this situation in the next time.By doing this, the artificial world is promoted as the parallel driving system run and the physical world is guided to be a more robust system.Overall, the approach has less computation time will be selected in actual experiments.Because the lower calculative time means the shorter time interval of TPM and indicates more comparison times of different control signals.Then,the controller could be more sensitive to the variation of this signal and more adaptive to the driving environments.Hence,we choose the EMD measurement priority in the real-time validated experiments in the future.

VI.Conclusion

Parallel driving is a promising solution to achieve high automation level.This paper introduces the parallel driving 3.0 system first, which consists of the control and management platform,intelligent vehicle platform and remote-control platform.The details and function of these three platforms are illustrated.To quantify the differences of the control commands between the artificial and physical world,several distance calculation methods are formulated for the probability distributions.Finally,the realization process and various applications of the parallel distance framework are concluded.

The presented approach aims to build a more general framework for distance comparison of arbitrary variables.The simulation results not only sheds light on comparison of signals,but also indicates the potential of the parallel distance framework in many other fields of automated driving,such as decision-making, parallel vision,parallel testing and parallel planning[36].Future work focuses on applying the parallel distance framework into the decision making of autonomous driving.For examples,to compare the differences of different control actions,such as acceleration,brake,lane change and lane keep.Parallel distance could help the artificial system tune the parameters quickly and ensure the algorithms converge to the optimal control actions.The hardware-in-theloop(HIL)and real vehicle experiments in parallel driving system can be leveraged to evaluate these controllers in the future.