Validation of Energy Saving From Cooperative Driving Automation via Vehicle-in-the-Loop Tests¹

Cooperative driving automation (CDA) involves technologies and systems that enable vehicles to communicate with each other as well as infrastructure and networks to make driving safer and more energy efficient. In contrast to conventional autonomous driving systems that depend mainly on internal sensors and computational algorithms, vehicles equipped with CDA, or connected and automated vehicles (CAVs), leverage communication technologies between vehicles (V2V) and between vehicles and infrastructure (V2I). This method promotes sharing information, collaborative decision-making, and coordination in real time, allowing vehicles to function in a unified and efficient way. The SAE J3216 Standard, proposed by Ref. [1], identifies four levels of CDA cooperation:

1. Status-sharing cooperation (class A)

2. Intent-sharing cooperation (class B)

3. Agreement-seeking cooperation (class C)

4. Prescriptive cooperation (class D)

At the status-sharing level, a CAV exchanges critical safety data with other participants, following the SAE J2735 Standard established by Ref. [2]. Intent-sharing allows a CAV to share its plans. For example, Correa et al. [3] proposed a system that shares the expected trajectories of the ego vehicle. Prescriptive cooperation takes a centralized approach, with all involved vehicles adhering to a collective strategy directed by a central coordinator. For example, the controlled platooning presented by Refs. [4,5] can be an example of prescriptive cooperation. However, its application in everyday passenger transport faces hurdles due to varying priorities and objectives among vehicle users.

Agreement-seeking cooperation offers a range of cooperation levels from individual to collective actions, blending elements from classes A, B, and D. In this class, an agent or coordinator proposes a cooperative action, and the other vehicles independently decide their level of participation based on negotiation and discussion, evaluating the advantages and consequences of joining the cooperative effort. This approach is relatively unexplored, with no standardized protocols currently established, highlighting the importance of further research into the mechanisms and designs that could optimize agreement-seeking cooperation. An extensive review of CDA planning and control strategies, including agreement-seeking cooperation, was conducted by Wang et al. [6].

Many studies have been conducted to quantify the energy impacts of CDA technologies through vehicle-in-the-loop (VIL) testing, mainly focusing on status-sharing cooperation. For example, Bae et al. [7] tested their energy-efficient cruising control strategies using status-sharing cooperation on both a dynamometer and roads. Energy-efficient approach and departure at signalized intersections using signal phase and timing messages have been tested on dynamometers [8] and on tracks ([9]). Additionally, Joa et al. [10] implemented an energy-efficient lane change strategy based on status-sharing cooperation on a track. For prescriptive cooperation, energy saving from platooning, especially for heavy-duty vehicles, has been extensively studied and tested for many years [11,12]. On the other hand, a few experimental studies have been conducted for intent-sharing and agreement-seeking cooperations. For instance, Wang et al. [13] tested intent-sharing cooperation in highway merging scenarios on a test track and assessed the benefits of safety and traffic efficiency benefits. Heß et al. [14] implemented cooperative lane changes entailing negotiation between two vehicles on a track and examined the significance of V2V communication performance for safety and efficiency. Given that CDA is an emerging technology in modern transportation, further studies are needed in these areas, particularly for realistic energy evaluation. Moreover, it is crucial to compare the performance of different CDA classes under the same testing conditions to reasonably compare the anticipated benefits of each class.

In our previous work, control systems for all the CDA classes, including a novel agreement-seeking cooperation system, were developed, and the control performance was evaluated under various conditions by Hyeon et al. [15]. This article extends our previous study by evaluating the energy benefits of the CDA controllers with an actual vehicle. First, the CDA control systems are embedded in the digital twin of the actual vehicle in CAV simulation software, and then the actual vehicle is connected with the digital twin and tested on a two-wheel chassis dynamometer. In this setup, the digital twin of the actual vehicle interacts with simulated environments and provides longitudinal control to the actual vehicle. Through this approach, the CDA engagement of the testing vehicle can be experimentally validated in scenarios involving multiple vehicles. Finally, vehicle data are collected and analyzed to verify the effectiveness of CDA systems in reducing energy consumption.

The main contribution of this article can be summarized as follows:

1. This work deployed different CDA classes, including our novel agreement-seeking cooperation system, in the VIL setup for car-following scenarios.

2. This work is the first attempt to compare the energy impacts of different CDA classes for the same scenarios and test conditions.

3. This article delivers the insights and possible limitations of implementing CDA, particularly needs to be considered in the operation of higher CDA classes.

The structure of this article is as follows: Sec. 2 provides a concise overview of the control system for each CDA class utilized in this work. Section 3 elaborates on the experiment setup and workflow, and Sec. 4 discusses experiment results. Finally, our conclusions are presented in Sec. 5.

The scenarios considered in this article are shown in Fig. 1. In all of the scenarios, three vehicles are driving in the same lane. The second and third vehicles, “CAV-1” and “CAV-2,” are CAVs equipped with CDA controllers. In this work, the CAVs transmit any information required for CDA operation with a 10 Hz frequency. The CDA control systems employed in this work were developed by Ref. [15]. The frontmost vehicle is not a CAV. This vehicle is included in the test scenarios to generate realistic driving traces for the following CAVs to follow. The method of determining the behaviors of the frontmost vehicle is elaborated in Sec. 3.2.

The rest of this section briefly reviews each CDA class’s systems, controllers, and scenarios validated in this work.

In status-sharing cooperation, the CAVs share their current position and speed, which are a part of basic safety messages defined by Cooperative Driving Automation Committee [2]. The vehicles individually plan their maneuvers based on the current status of the surrounding vehicles. The controller used in this work, named eco-adaptive cruise control (eco-ACC) by Ref. [15], is a model predictive control (MPC) producing energy-conscious car-following behaviors. The optimal control problem (OCP) is formulated in a quadratic program, minimizing control efforts and ensuring a safe distance from a preceding vehicle. Therefore, the CAVs require the preceding vehicle’s future maneuvers over an MPC prediction horizon to avoid a collision. For this purpose, the speed forecasting algorithm developed by Hyeon et al. [16] is used in this work, which uses polynomial regression to predict the preceding vehicle’s future speed over a short-term future. The prediction horizon of this work is set to 3 s. More information on control formulation and parameter settings can be found in the article by Hyeon et al. [15].

Intent-sharing cooperation is similar to status sharing in that vehicles plan their maneuvers separately using the eco-ACC. The difference from status sharing is that the CAVs share future intentions. In this work, the future intentions include the position and speed trajectories over a prediction horizon of 3 s. Using a predictor may not be necessary during intent-sharing unless the CAV’s MPC prediction horizon length exceeds the time horizon of the intent shared by other vehicles. This work assumes that the CAV’s MPC and the intent-sharing messages have the same preview length of 3 s, and therefore, the predictor is not used in intent-sharing cooperation unless the preceding vehicle is driven by a human. Consequently, the CAV-1 needs to use the predictor because its preceding vehicle is human driven and does not share its intents.

Prescriptive cooperation, within the framework of CDA, is the highest level of cooperation. This class is characterized by a centralized approach to coordination in which a central authority or system dictates specific actions for the participating CAVs. In this work, CAV-1 in Fig. 1 is designated the central coordinator that plans the platooning operation and sends a proposed trajectory to CAV-2, and CAV-2 obeys CAV-1’s plan for the entire trip. A cooperative driving plan, or platooning trajectory, is computed based on the OCP proposed by Hyeon et al. [15]. This controller, named eco-platooning control in our previous article, solves an OCP similar to the eco-ACC but optimizes for the group of vehicles. Therefore, the solution of the eco-platooning control OCP is energy-conscious platooning trajectories for all the platooning vehicles. The trajectories include the distance from the immediately preceding vehicle, speed, and acceleration over a 3 s future. In addition, CAV-2 sends its current status to CAV-1, where it is used for CAV-1’s planning.

The objective of agreement seeking in this work is to form a platoon with CAV-1 and CAV-2. The agreement-seeking process is visualized in a simple diagram in Fig. 1. In our scenarios, the leading CAV, referred to as the CDA initiator, initiates the process by sending cooperative operation requests to the following CAVs. The CAVs that receive these requests are termed CDA recipients. CAV-1 is the CDA initiator in our testing scenarios, and CAV-2 is the CDA recipient, as noted in Fig. 1. The requests dispatched by CAV-1 include proposed trajectories such as position, velocity, and acceleration. Upon receiving the request, CAV-2 assesses the suggested plan, decides whether to agree with the proposal, and communicates its decisions back to the CDA initiator. The rest of this section elaborates on each phase of this process.

Initially, CAV-1 invites CAV-2 to form a platoon, acting as the CDA initiator. At this stage, CAV-1 proposes a trajectory for CAV-2 to follow upon agreement. This trajectory, detailing the distance gap, speed, and acceleration for the next 3 s, is calculated using the eco-platooning controller described in Sec. 2.3.

Upon receiving the cooperative driving plan, CAV-2, now becoming the CDA recipient, evaluates the proposal using its decision-making logic. The CDA recipient decides on acceptance or rejection based on the evaluation results. Various decision-making techniques can be applied here, from simple heuristics to advanced multi-agent reinforcement learning [17]. In this work, the simple heuristic method developed is employed to minimize computational loads. In this method, CAV-2 compares the cost function values from the cooperative plan shared by CAV-1 and its individual plan. The mathematical derivation and detailed information can be found in the study by Hyeon et al. [17]. The CDA recipient then returns its decision to the CDA initiator.

The CDA initiator decodes the decision sent by the CDA recipients to discover whether a platoon formation is feasible. The outcome will be one of the two cases:

1. If the CDA recipient rejects the proposal, a platoon cannot be formed, and both CAV-1 and CAV-2 will employ their own eco-ACCs. Then, the CAV-1 will become a CDA initiator again and start another proposal, as described in Sec. 2.4.1.

2. If the CDA recipient agrees to the proposal, a platoon can be formed. The CDA initiator then calculates the final cooperative plan for the next 3 s using the eco-platooning controller. This plan is then transmitted to the CDA recipient. The CDA recipient adheres to the shared plan unless it becomes outdated. This cooperative mode remains in effect for a predetermined period, chosen as 30 s in this work. Once this period ends, the CAV-1 automatically reinitiates the proposal, as explained in Sec. 2.4.1.

The VIL testing setup is depicted in the schematic in Fig. 2. This setup is composed of three parts: (1) the simulation rendering virtual environment and the digital twin of the test vehicle, (2) the test vehicle equipped with instruments for recording data and connecting the simulation environment, and (3) the testing facility. The following sections describe the details of our testing setup and introduce the baseline model and drive cycles selected for this study.

To evaluate the energy benefits of implementing CDA technologies, we need to compare their energy consumption to that of vehicles driven by humans. Using an accurate human driver model as a baseline is crucial because it can significantly affect the estimation of energy savings impacts. For this purpose, we adopted a human driver model (HDM) developed by Han et al. [18]. This model is designed to replicate the car-following behaviors of a human driver and was thoroughly validated by actual human driving data, showing higher accuracy compared to the intelligent driver model. The model parameters can be categorized by situation: free flow and car following. As the scenarios in our study involve car-following situations, the essential parameters include minimum gap distance at standstill conditions (⁠$dmin$⁠) and the time headway from the preceding vehicle (⁠$τdes$⁠). The values are selected as $dmin=5m$ and $τdes=1s$⁠, which are the same values of the CDA controller’s parameters for fair comparison.

The simulation setup is required to render the digital twin of the test vehicle in the virtual environment. In this work, three vehicles drive in a row, as shown in Fig. 2. While the front two vehicles virtually exist in simulation, the third vehicle is a digital twin, mirroring the actual vehicle’s behavior. This digital twin is equipped with the control modules designed to generate the control commands necessary to guide the actual vehicle.

The frontmost vehicle, labeled “Dummy” in Fig. 2, follows predetermined drive cycles to simulate downstream traffic conditions encountered in real-world driving. This study employs two drive cycles developed by the U.S. Environmental Protection Agency (EPA): the Urban Dynamometer Driving Schedule (UDDS) and US06. These cycles have been commonly utilized to assess vehicle energy consumption: UDDS represents a light-load, low-speed cycle averaging around 30 mph, and US06 is characterized by its more aggressive acceleration and braking, entailing higher speeds and loads, and resulting in increased energy consumption. The second and third vehicles are governed by either the HDM or CDA controller.

The simulation software roadrunner, developed by Argonne National Laboratory (Argonne), is used to construct a virtual environment and digital twin of the actual vehicle. roadrunner enables the construction and simulation of both CAVs and non-CAVs in a dynamic environment. This platform supports the integration of high-fidelity powertrain models, allowing for precise assessments of vehicle energy consumption under varied conditions. For a comprehensive overview of roadrunner’s features and capabilities [19].

In this work, a midsize electric vehicle model is used to simulate the dummy vehicle and CAV-1. For CAV-2, we used the 2017 Prius Prime model extracted from Autonomie developed by Jeong et al. [20]. During the tests, the CAVs can interact with the surrounding vehicles by perceiving them through onboard sensors and V2V communication models. In this work, all the communication required for CDA operations is simulated in roadrunner.

The experiments outlined in this article were conducted at Argonne using a two-wheel-drive chassis dynamometer. This dynamometer can accommodate vehicles ranging from light- to medium-duty, supporting up to 223 kW (300 hp) at the axle. The temperature of the testing environment is controlled to maintain $72∘F$ and equipped with a continuous-speed fan to ensure necessary cooling. For this study, the 2017 Prius Prime, equipped with an advanced package, was selected as the vehicle for experimentation. A photo of the test vehicle mounted on the dynamometer is shown in Fig. 3. For more information on the test vehicle, refer to the work of Di Russo et al. [21].

To ensure consistency across the tests, we managed the initial conditions of transmission oil temperature and the battery state of charge (SOC) to be in a similar range in each test. After every run of the test, the vehicle was recharged to 85% of SOC while maintaining the powertrain temperature level. All the tests presented in this article were conducted in full electric vehicle mode to avoid any interference from energy management.

The simulation is run on a desktop computer and exchanges information in real time with the vehicle and with an onboard controller installed in a vehicle (dSPACE MicroAutoboxII) using the Ethernet UDP protocol. The onboard controller, responding to control commands from roadrunner, activates the vehicle’s actuators and feeds the vehicle’s state back to the simulation. This testing architecture streamlines the testing process by separating vehicle-specific and vehicle-agnostic components, allowing for rapid and automated execution of a large number of scenarios and real-time calibrations.

The vehicle is equipped with a Hioki PW8001-10 power analyzer, which directly measures voltage and current for energy consumption analysis. Additional data on vehicle and component operation were acquired from data exchanged over the vehicle controller area network (CAN) bus. CAN bus messages were recorded by neoVI Fire3 from Intrepid Control Systems Inc., which enables vehicle interface, signal gateway, and data logging. To ensure precise analysis, a customized data acquisition system at the Argonne facility collects and merges data from roadrunner and all instruments, producing time-aligned datasets in a single file.

To evaluate the precision of our results, we measured the differences between the control demand signals and the vehicle’s actual performance using root-mean-squared error (RMSE). The RMSE of tracking errors is computed by sampling errors from the test results with all types of controllers. The results show that the RMSE of speed demand tracking is 0.09 m/s and 0.08 m/s in the US06 and UDDS cases, respectively. Figure 4 presents one example of actual and demanded trajectories in agreement-seeking cooperation in the US06 scenario. It includes detailed zoom-ins on speed and acceleration between 500 s and 550 s, confirming the vehicle’s ability to follow the control demands throughout the trip precisely. The acceleration demand and the actual acceleration may not match perfectly due to the adjustment made by the vehicle’s controller, which modifies the acceleration command based on various factors such as transmission gear, torque, and speed. In contrast, the speed demand is calculated from the current speed of the vehicle, resulting in a closer match between the demanded and actual speeds.

In order to confirm the reliability of the VIL test outcomes, the driving performance of CAV-2 in a pure simulation environment was compared with the results of the experiments. The simulation study uses the 2017 Prius Prime vehicle model with a high-fidelity powertrain model for CAV-2. The same virtual environment and scenarios are used in both simulation and experiment studies. This section compares vehicle states and agreement seeking outcomes from simulation and experimental studies.

The differences between simulation and experiment results are quantified using RMSE, relative RMSE (RRMSE), and the maximum magnitudes of errors, listed in Table 1. As with the RMSE of tracking, the values in Table 1 are computed by combining all the controllers’ results for each drive cycle scenario. The results show that the RMSE of speed is less than 0.5 m/s, verifying that CAV-2’s driving behaviors in simulations and experiments are sufficiently close.

Measuring CR is important for our analysis for two reasons: First, in the experiments, decision-making results can be affected by signal delays or turnaround time. Demonstrating a match between the simulation and experiment CR results proves that the signal turnaround time is sufficiently short for dynamic decision-making. Second, total cooperation time is a key factor determining energy savings, as discovered by Hyeon et al. [17]. Therefore, accurate energy evaluation is more likely with close CR results from simulations and experiments.

In Table 2, the CR from the simulations and experiments of agreement-seeking cooperation are compared. The results indicate that the CRs computed from simulation results and experiment results have comparable values: 0.03 % and 6% difference in UDDS and US06, respectively. Overall, the differences between simulation and experimental results, including CR and trajectories, are larger in the US06 scenario than in the UDDS. One reason is the high-speed and high-acceleration driving patterns inherent in the US06 scenario. Such driving patterns are more susceptible to delays during synchronization between the digital twin and the actual vehicle.

The energy consumption of CAV-1 and CAV-2 was compared with that of HDM. The bar graphs in Fig. 6 show the energy consumption compared to the HDM when CAVs execute controllers for different CDA classes. Here, the CAV-1 is virtual, and the results are obtained from the roadrunner powertrain model. CAV-2 is the actual vehicle, and its energy consumption was recorded using the Argonne facility. Figures 6(a) and 6(b) present the results of CAV-1 and CAV-2, respectively, and their average values are shown in Fig. 6(c). In the results, a trend can be found that the CDA’s energy-saving impact is more significant in US06 than in UDDS. This observation was to be expected because US06 is a more aggressive drive cycle and thus provides a larger potential for energy improvement. In addition, the energy saving is greater in CAV-2 than in CAV-1 because the eco-ACC and eco-platooning control reduce the fluctuating speed often occurring in upstream traffic, which demands more energy consumption.

It is noteworthy that even cooperation with only two vehicles achieves an energy saving of 5% in the actual vehicle with high cooperation levels (class C and D). Since energy saving grows along the platoon end, we believe that the observed energy saving can grow with more CAVs under the high penetration of CDA class C and D. Finally, we would like to spotlight the performance of agreement-seeking cooperation (class C). In both simulation and experimental results, agreement-seeking cooperation offers marginally lower energy-saving performance than prescriptive cooperation; for instance, with the agreement-seeking cooperation, the average energy saving of CAV-2 is only 0.3% lower than that of the prescriptive. Considering that agreement-seeking cooperation is a more feasible technology for general passenger vehicles, the importance of further research on agreement-seeking cooperation is clear.

This article validates energy savings achieved through various classes of CDA by employing VIL testing. The VIL experiments were conducted on a chassis dynamometer with various CDA controllers and scenarios by operating the digital twin of the 2017 Prius Prime in simulation. In the experiment, agreement-seeking cooperation offered higher energy savings than a lower cooperation level and, in fact, as much as prescriptive, the highest level of cooperation. Although prescriptive cooperation can bring more energy savings, continued research in agreement-seeking cooperation is crucial given the practical applicability of agreement-seeking cooperation for mainstream passenger vehicles. In future work, we plan to test our CDA controllers with realistic V2V communication by integrating V2V onboard units into the VIL testing setup. Finally, a comprehensive analysis considering the safety and traffic efficiency of CDA operation would also be valuable.

The submitted manuscript has been created by UChicago Argonne, LLC, Operator of Argonne National Laboratory (“Argonne”). Argonne, a U.S. Department of Energy Office of Science laboratory, is operated under Contract No. DEAC02-06CH11357. The Department of Energy (DOE) will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan.

There are no conflicts of interest.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.