Cyclops: Open Platform for Scale Truck Platooning

Fig. 3 shows the overall architecture of our testbed, where three trucks (LV, FV1, and FV2) are connected through a wireless network to each other and also with a remote control center. The control center communicates directly with LV, providing the target velocity and gap distance. The target velocity is used as the reference velocity of LV, while the target gap distance is communicated to FV1, which is relayed to FV2 for velocity and gap control. We point out that every truck in Cyclops is autonomous, which is not valid in full-size truck platooning, where its LV is usually human-driven.

In FV1 and FV2, the perception system receives the front-facing camera images and the lidar data, where the camera image is used for detecting the lane curvature, while the lidar points are used to locate the preceding trailer’s rear end. Then the following information is delivered to the control system: (i) the lane curvature, (ii) the gap distance, (iii) the current velocity from the encoder sensor, (iv) the target gap distance given by the command center, and (v) the preceding truck’s reference velocity. Now the lane keeping controller decides the steering angle, controlled by the steering motor through the steering motor controller. The gap controller generates the reference velocity to meet the target gap, then given to the velocity controller that controls the driving motor through the driving motor controller. The reference velocity generated by the gap controller is communicated to the following truck through V2V communication for its gap control.

LV and FVs share the same hardware and software architecture with a slight difference. Since LV does not need the gap controller, it is replaced with the emergency braking controller between the obstacle detection and the velocity controller. Since our testbed does not allow any other surrounding vehicles except the trucks at this moment, LV automatically enforces an emergency stop when it detects an obstacle in the driving direction, which in turn triggers emergency stops in FV1 and FV2. Additionally, the command center can also enforce an emergency stop by a remote command.

Fig. 4 shows one of our 1/14 scale semi-trailer trucks. As the truck body, we use an RC truck famous among the RC enthusiasts. Three such trucks were developed, one designated as LV while the others are FVs. The figure highlights nine essential components such as sensors and computing platforms. The hardware components are listed in Table I. More specifically, each truck is equipped with a 180 degrees FoV (Field of View) front-facing camera on top of the tractor. A single-plane lidar sensor with a 25 m detection range is hidden inside the truck cabin. An encoder that counts wheel rotations is installed at the tractor’s front wheel to measure the vehicle velocity. We use two heterogeneous computing platforms for the high-level and low-level control, respectively. For the high-level controller, we use an Nvidia Jetson AGX Xavier located inside the trailer, while, as the low-level controller, an OpenCR embedded computer is attached at the back of the tractor. For the driving motor, we use a 1800 KV BLDC motor. For the steering motor, we use a servo motor with a torque of 3.3 kg$\cdot$cm. All the electronic components are powered by a 96 Wh battery in the trailer. We use the WIFI network interface in the Nvidia Jetson AGX Xavier for wireless communication, which is reinforced by an additional high-performance antenna installed on the trailer.

The objective of lane detection is to find the curvature of the desired path from the front-facing camera. For that, a camera image goes through the following steps: (i) pre-processing of cutting the trapezoid ROI and removing noises using the Gaussian blur method; (ii) conversion into a binarized bird’s eye view image; (iii) sliding window-based search to detect the left and right lanes; (iv) fitting the desired path’s curvature combining the left and right lane curvatures. The final curvature is decided as a second-order polynomial.

The above lane detection method has been widely used in many studies [10, 11, 12] because it is robust and lightweight for the implementation in small embedded systems. Unfortunately, however, the method does not perform well in the platooning scenario due to the minimal gap to the preceding truck. Fig. 5(a) shows a forward view image and its lane detection result with the static ROI method where the trapezoid ROI is a fixed one. The binary image indicates that the slight inclusion of the trailer’s rear end in the ROI badly affects the lane detection result. The noise by the preceding trailer is a known issue reported in the platooning with full-size trucks [9]. In our testbed, the effect is even vibrant because the trailer’s color happens to be white.

To eliminate the noise by the trailer at a close distance, we try to cut off the trailer portion in the bird’s eye view image. For that, we use the gap information from the lidar sensor. By actual measurements, we found that the ratio between the longitudinal distance and the number of vertical pixels is 490 pixels per 1.0 m. With this mapping, Fig. 5(b) shows the adapted ROI that precisely removes the noise, which provides a significantly improved lane detection result. This adaptation logic is added in our lane detection module such that it can dynamically determine the optimal ROI.

Obstacle detection provides the gap distance to the preceding truck. We use a single-plane lidar sensor that provides a set of points to the nearest objects around 360 degrees. To detect the preceding truck, we limitedly use the front-facing 60 degrees. For the obstacle detection, we employ the well-known obstacle_detector ROS package [13], which detects and also tracks nearby objects by clustering the 2D lidar points. The tracking result is obtained as a circle where its center and diameter correspond to the middle point and the horizontal length of the preceding trailer’s rear end, respectively.

In this section, we introduce the velocity control, which consists of the feed-forward plus anti-windup proportional-integral (FF+AWPI) control, the velocity saturation function ${\rm sat}_{i}(u_{v,c,i})$, and the inverse of the nonlinear function $f_{i}^{-1}(\bar{u}_{v,c,i})$. Here, the subscript $i=0$ denotes the variables or function for LV, $i=1$ for FV1, and $i=2$ for FV2. The FF+AWPI control calculates the velocity control input $u_{v,c,i}$ such that

$$\begin{split}u_{v,c,i}(k)&=K_{F,i}v_{r,i}(k)+K_{P,i}\tilde{v}_{i}(k)+\sum_{j=0}^{j=k}{K_{I,i}\tilde{v}_{i}(j)T_{s}}\\ &\ \ \ +\sum_{j=0}^{j=k-1}{K_{A,i}\left(\bar{u}_{v,c,i}(j)-u_{v,c,i}(j)\right)},\end{split}$$

(1)

where $T_{s}$ is the sampling time, $\bar{u}_{v,c,i}$ is the velocity control input saturated by function ${\rm sat}_{i}(u_{v,c,i})$, the positive constants $K_{P,i},K_{F,i},K_{I,i}$, and $K_{A,i}$ are control parameters, and $\bar{v}_{r,i}$ is the saturated velocity reference. The velocity tracking error $\tilde{v}_{i}$ is given as

$$\begin{split}\tilde{v}_{i}(k)=\bar{v}_{r,i}(k)-v_{i}(k),\end{split}$$

(2)

where $v_{i}$ is the velocity measurements of the truck.

The saturation function ${\rm sat}_{i}(u_{v,c,i})$ represents the velocity limitation that a truck can generate. The saturation function ${\rm sat}_{i}(u_{v,c,i})$ is given as

$$\begin{split}{\rm sat}_{i}(u_{v,c,i})=\begin{cases}V_{i}^{\max},&{\rm if}\ u_{v,c,i}\geq V_{i}^{\max}\\ 0,&{\rm if}\ u_{v,c,i}\leq 0\\ u_{v,c,i},&{\rm otherwise}\end{cases},\end{split}$$

(3)

where $V_{i}^{\max}$ denotes the maximum velocity that a truck can generate.

The motor command and the velocity of the truck have the nonlinear relation. The inverse of nonlinear function $f_{i}^{-1}(\bar{u}_{v,c,i})$ is used for linearly treating the truck. The nonlinear function $f_{i}(u_{i})$ representing the nonlinearity between the motor command and the velocity of the truck is obtained by fitting the experimental data for the motor command and the velocity of truck into a quadratic function, and is given by

$$\begin{split}f_{i}(u_{i})=a_{i}u_{i}^{2}+b_{i}u_{i}+c_{i},\end{split}$$

(4)

where $a_{i},b_{i}$, and $c_{i}$ are design parameters. From (4), we can induce the function $f_{i}^{-1}(\bar{u}_{v,c,i})$ such that

$$\begin{split}f_{i}^{-1}(\bar{u}_{v,c,i})=\frac{-b_{i}+\sqrt{b_{i}^{2}-4a_{i}(c_{i}-\bar{u}_{v,c,i})}}{2a_{i}}.\end{split}$$

(5)

Finally, the output of the function $f_{i}^{-1}(\bar{u}_{v,c,i})$ is the motor command $u_{i}$, which is applied to the driving motor.

The gap control is used by only FVs. The goal of the gap control is to make the gap to the preceding truck closely follow the gap reference. The gap control is composed of the feed-forward plus proportional-differential (FF+PD) control and the velocity reference saturation function ${\rm sat}_{r,i}(v_{r,i})$. The FF+PD control is given in the following form:

$$\begin{split}v_{r,i}(k)&=\bar{v}_{r,i-1}(k)-K_{GP,i}\tilde{d}_{i}(k)\\ &\ \ \ -\frac{K_{GD,i}(\tilde{d}_{i}(k)-\tilde{d}_{i}(k-1))}{T_{s}},\end{split}$$

(6)

where $\bar{v}_{r,i-1}$ is the saturated velocity reference of the preceding truck received via wireless communication. The positive constants $K_{GP,i}$ and $K_{GD,i}$ are control parameters, and the gap tracking error $\tilde{d}_{i}$ is obtained as

$$\begin{split}\tilde{d}_{i}(k)=d_{r,i}(k)-d_{i}(k),\end{split}$$

(7)

where $d_{r,i}$ denotes the gap reference and $d_{i}$ denotes the measured gap to the preceding truck.

The velocity reference saturation function ${\rm sat}_{r,i}(v_{r,i})$ saturates the velocity reference as the velocity limit of the lane, which is represented as

$$\begin{split}{\rm sat}_{i}(v_{r,i})=\begin{cases}V_{r}^{\max},&{\rm if}\ v_{r,i}\geq V_{r}^{\max}\\ 0,&{\rm if}\ v_{r,i}\leq 0\\ v_{r,i},&{\rm otherwise}\end{cases},\end{split}$$

(8)

where $V_{r}^{\max}$ denotes the velocity limit of the lane. Finally, the saturated velocity reference $\bar{v}_{r,i}$ plays as the velocity reference for the velocity control.

In this section, we introduce the lane keeping control to maintain the truck within the lane. The lane keeping control is given by

$$\begin{split}\delta_{i}(k)=-K_{i}(v_{i})e_{i}(k)-K_{L,i}(v_{i})e_{L,i}(k),\end{split}$$

(9)

where $\delta_{i}$ is the steering angle, $e_{i}$ is the estimated lateral error, $e_{L,i}$ is the preview lateral error, and $K_{i}(v_{i})$ and $K_{L,i}(v_{i})$ are control parameters that are varied according to the truck velocity. Because faster velocities lead to bigger lateral changes by a given steering angle, with a faster velocity, the control parameters have to be smaller.

In (9), the preview lateral error $e_{L,i}$ is obtained with a camera sensor. The lateral error $e_{i}$, however, cannot be directly obtained with the camera sensor due to its limited sight, thus we estimate the lateral error $e_{i}$ such that

$$\begin{split}e_{i}(k)=e_{L,i}(k)-L_{i}{\rm tan}(\theta_{i}(k)),\end{split}$$

(10)

where $L_{i}$ is the preview distance and $\theta_{i}$ is the angle between the direction in which the truck moves and the preview point on the desired path which is obtained with the camera sensor.

Fig. 7 shows the evaluation results for the first platooning scenario (Scenario 1). Fig. 7(a) shows how LV follows the given reference velocity as it ramps up from 0.6 m/s to 1.2 m/s. Although the results show slight overshoots upon velocity changes, they do not last long, resulting in satisfactory longitudinal control performance. Fig. 7(b) shows the measured gaps between trucks as the reference gap is constant at 1.2 m. The results show that all the trucks maintain the desired gap well despite the acceleration of the platoon. The lateral errors, meanwhile, increase upon the velocity goes about 1 m/s, but all the trucks are safely maintained within the lane.

In Fig. 8, the second scenario (Scenario 2) is validated, where the reference gap ramps down from 1.2 m to 0.6 m while maintaining a constant velocity at 1.0 m/s. Fig. 8(a) shows the LV velocity well maintained at 1.0 m/s. Fig. 8(b) shows that the measured gaps stick well to the dynamic reference gap. However, in Fig. 8(c), we can see that the lateral errors get significant while the reference gap is the smallest (i.e., 0.6 m). As mentioned before, this lateral safety problem happens when the gap is too short such that the camera’s view is significantly occluded by the preceding trailer.

We apply the dynamic ROI method to the lane detection module for solving the lateral stability problem. Fig. 9 compares FV2’s lateral errors with static ROI and dynamic ROI while maintaining the 0.6 m gap between FV1 and FV2. In the figure, the lateral error with dynamic ROI is visibly smaller than that with static ROI. Even worse, FV2 goes out the lane at the 51.4-second point when operated with the static ROI method. In contrast, the truck is safely maintained within the lane with dynamic ROI. Thus, we claim that dynamic ROI is an effective method to enhance lateral stability and safety.