IEEE Robotics & Automation Magazine - September 2023 - 45

compared to results obtained when using constant parameters
to identify their respective strengths and weaknesses.
They have been implemented and tested in real conditions
on an off-road mobile robot with varying terrain and trajectories.
An in-depth analysis of the proposed techniques is
done, and further insights are obtained in the context of
gain tuning for steering controllers in dynamic environments.
The performance and transferability of these methods
are demonstrated as well as their robustness to changes
in terrain properties. As a result, tracking errors are reduced
while preserving the stability and explainability of the control
architecture.
INTRODUCTION
The problem of tuning controller gains is crucial in robotics as
it defines the performances and the behavior of robots. As is
well known, it relies on low-level performances as well as
interaction conditions between robots and their environments.
It is especially the case in autonomous navigation of off-road
mobile robots as achievable performances (settling time or settling
distance of the robot, i.e., the time required to reach a
desired setpoint within a set margin) depend not only on actuator
delays but also on grip conditions and velocity. In the
framework of path tracking problems, the use of constant gains
for a robot able to move at different speeds, on several types of
grounds, and with complex trajectories leads to suboptimal
results with potentially important tracking errors. This is a
drawback, as accuracy is often crucial in off-road applications,
especially in agriculture, where a few centimeters of accuracy
is targeted, and robots may move on narrow paths, bordered by
a ditch.
Consequently, the use of an online gain adaptation arises as
an interesting approach to optimize robot behavior. Some methods
for gain tuning are mainly developed in the framework of
motor control (such as in [1], utilizing neural networks, or in [2],
using a performance index). In the framework of mobile robotics,
the adaptation is mainly focused on model parameter estimation
(such as in [3]) but is rarely devoted to gain tuning. However, some
recent works [4], [5] show that online gain tuning of
a path tracking control law may improve navigation
stability. An optimal solution to gain tuning is difficult
to obtain formally. A robot's behavior indeed
depends on many factors (e.g., grip conditions,
localization uncertainty, and noises), which can
be difficult to measure or estimate and are, moreover,
variable in off-road contexts [6]. As a result,
the online variation of control parameters has to be
done carefully as it can lead to unstable behavior.
In this article, an adaptive and predictive control
strategy, developed previously in [7] with constant
parameters, is used to achieve an accurate path
tracking. Therefore, an observation strategy [8] is
available to access the cornering stiffness, together
with the residual of a Kalman filter approach for
localization [9]. This allows for the explicit access
of information related to grip conditions and confidence
in the localization process to set online the
gains of an adaptive and predictive control law for
path tracking. The first method, based on a dynamic
model of the robot [10], is suggested to define an
explicit relationship between the control parameters
and the robot's state (velocity, contact properties,
and so on). Based on the estimation of the yaw rate
settling time, this first approach shows interesting
results but does not account for unexpected or nonmodeled
deviations, especially related to localization
accuracy.
Accordingly, the second approach, based on a
neural network [11], is offered in this article and
categorized as policy iteration reinforcement
learning (RL) [12]. This allows, with the help of
a training simulator, for optimization of a function,
accounting for both deviation and potential
oscillations. This way, control law gains may be
modified online to reduce the path tracking error
SEPTEMBER 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
45
IMAGE LICENSED BY INGRAM PUBLISHING
IEEE Robotics & Automation Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2023
