IEEE Robotics & Automation Magazine - December 2022 - 101

used to switch between the simulation and real-world environments.
For running both the simulation and real-world
environments, the compact observations are the joint positions
.zt The compact observations are first fed into the informative
network to predict the full observations
o .i
t
Then, the
NN-based controller outputs the actions to be executed either
in the simulation or the real world. For real-world experiments,
the NN-based controller is computed on a desktop
computer and the actions are sent to an Arduino DUE via
Robotic Operating System (ROS) messages. Then, the Arduino
DUE serves as a host to transfer the ROS messages into
CAN messages and sends them to the STM32 controller on
each module via a CAN bus.
Experiment Setup
The power consumption can be calculated using its running
current and voltage. In our experiments, we use a 24-V constant
voltage power supply to actuate the robot. The current
can be acquired by measuring the voltage of a high-load resistor
(/ .),
power consumption of the prototype snake-like robot can be
calculated as
()
pp 24 /( ),
.
U 24 V . The voltage Up
scope, which samples at 1,000 Hz.
p =
is measured using an oscilloResults
Due
to time and hardware durability, we tested only five target
velocities for the NN-based controller in our prototype
experiments, namely, 0.05, 0.10, 0.15, 0.2, and 0.25 m/s. We
controlled the robot to run for 10 s to measure power consumption
and actual velocity. The actual velocities for the
gaits generated by the NN-based controller were 0.03, 0.05,
0.08, 0.12, 0.15 m/s, approximately, which were all less than
the target velocities. This is because of the reality gap between
the simulated environment and the real world. For the same
reason, the performance of the gait equation controller is also
impacted by this reality gap. To fairly compare energy efficiency,
we chose parameters that lead to gaits traveling at
these five actual velocities with the best energy efficiency in
the simulation.
First, we present montages of the slithering gaits at 0.15 m/s
50 W0 1X which is stringed into the circuit. Thus,
PU RUp=- where R 01p = X and
from the gait equation controller and the NN-based controller
(see Figure 7). To show a full cycle of the movement, seven
video frames (25 frames per second) are selected to visualize
the pattern of the gaits generated by both controllers. For the
gait generated by the gait equation controller, frames 89-119
show one full cycle of the gait. For the gait learned by the NNbased
controller, roughly two motion cycles are shown. As we
can see, the body curve of the gait equation gait is shaped as a
sinusoid wave. The body curve of the RL gait is slightly more
slender than the body curve of the parameterized gait. This
slider-body curve can propel the robot more smoothly and
avoid joint power consumption as much as possible, which is
in accordance with the simulation results. We can also
observe that the learned gait moves at a higher frequency
than the parameterized gait and moves straight toward the
forward direction, while the parameterized gait tends to deviate
from the forward direction.
Second, we show results of the power consumption of the
gaits from both the gait equation controller and the NNbased
controller, which are presented in Figure 8(a). All of
the power consumption data were measured in 10 s for fair
comparison. The solid line represents the exponentially
weighted moving average of the power consumption, and the
error bar depicts standard deviation of the raw power consumption
IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022
