IEEE Electrification Magazine - March 2020 - 45

offers the possibility to optimize the efficiency of the power-transfer process for all of the possible loading and coupling conditions. The situation in which the sending and
pickup converters are operated in voltage-regulation
mode to achieve maximum control flexibility is depicted
in Figure 5(b). Note that the active H-bridge implemented
onboard the truck prototype, as shown in Figure 4, gives
full control over the load impedance (active and reactive parts) within the capabilities of the converter, thus
enabling the fine tuning of the pickup resonant frequency.
The H-bridge converter can also be operated as a diode
rectifier, making it possible to investigate a wide range of
control methods without redesigning the hardware.

Implemented Functions for Autonomous
Operation
Two methods for self-driving operation were implemented and tested with the truck model: a simultaneous
localization and mapping (SLAM)-based method with a
path-tracking algorithm and a supervised machinelearning method. For the SLAM-based path tracking,
Hector SLAM, developed by Team Hector from the Technische Universität Darmstadt, was chosen as the preferred method. Hector SLAM represents an odometry-free
SLAM solution for ROS and was used to obtain a map of
the environment and estimate the pose of the vehicle.
For simple demonstration purposes, a waypoint-logging
node was written to manually record a reference path for
the truck model to follow. This is combined with the popular Pure Pursuit steering controller to enable autonomous path tracking. Pure Pursuit receives the recorded
path in the form of a list of waypoints through the ROS
network. It then calculates the appropriate curvature
with a corresponding steering angle so that the vehicle
can move from its current position to a look-ahead goal

point on the reference path. The forward velocity is set to
a fixed value.
As the vehicle moves forward, it pushes the goal point
forward on the path, with a predefined look-ahead distance. For the implementation, the tracking of the driving
path depends only on the lidar, while the camera is mainly used for observation. The implementation enables the
truck model to autonomously follow the recorded path,
operating in a closed loop. As a result, this function supports a very convenient demonstration, where the autonomous model can be left to operate continuously on a path
that includes the dynamic wireless-charging section. If
sufficiently long roadside coil sections were introduced in
the path to ensure an average energy balance, the truck
model could be left to operate continuously on a defined
track, with a theoretically infinite driving range.
As an example of autonomous operation, Figure 6 presents a case where the initial reference path recorded by the
lidar is shown in red, while the trajectory of the truck
when it autonomously tracks the path through five laps is
shown in blue. As indicated by the figure, the path-tracking
algorithm has a reasonably accurate and consistent performance, and the system can be left to operate continuously.
A CNN inspired by the NVIDIA DAVE-2 self-driving car
has also been implemented for the truck model, with
Keras as a separate mode to enable camera-based autonomous driving. This is useful for achieving autonomous
driving in large, open areas where the lidar is out of range.
The network was trained from camera images and steering-angle data from manually driving the truck on a visible track marked on the floor. When the training process is
complete, the truck can drive itself on the track by outputting a steering angle from the corresponding incoming
camera images. For brevity, no explicit results are shown
here, but a performance similar to the SLAM-based path

Figure 6. An example of the autonomous-truck model's position trajectory when it is operated with path tracking.

IEEE Elec trific ation Magazine / MARCH 2 0 2 0

IEEE Electrification Magazine - March 2020

Table of Contents for the Digital Edition of IEEE Electrification Magazine - March 2020