IEEE Robotics & Automation Magazine - December 2022 - 102

observe that the power consumption increases with movement
velocity, which ranges from approximately 21.5 to 28 W.
As the velocity of the gait generated from the gait equation
controller is directly related to the frequency of the sinusoid
wave, it moves much slower at a low speed and faster at a
high speed, thereby consuming more power. We can see that
the NN-based controller consumes more power at a low
speed. This is because the controller generates a concertina
gait, which only moves parts of the body to inch forward and
those static joints require the locked-motor current, thereby
consuming more power. Another potential reason is that the
NN-based controller has difficulties maintaining low and
accurate velocities. The small adjustments of each joint will
impact the overall velocity observably and thus consumes
more energy. Compared to the power consumption of the
gait equation controller, we find that the NN-based controller
can increase energy efficiency by 5 to 10%, especially at
intermediate and high velocities. For both controllers at one
velocity, power consumption shows an obvious fluctuation,
which can be interpreted as the periodic movement of the
robot. Figure 8(b) depicts the power-profile comparison of
the simulation and real-world results. We can find that there
is still a performance gap between the simulation and realworld
experiment due to some potential reasons. For
instance, it is challenging to simulate the environment with
perfect fidelity, which therefore leads to the performance
drop. Inaccurate sensory measurements or noisy observation
stats can also cause the performance drop.
Gait Equation
Controller
Frame 89
Frame 94
Frame 99
Frame 104
Frame 109
Frame 114
Frame 119
Direction
0.6 m
Figure 7. Montages of the snake-like robot performing slither gait
at 0.15 m/s, controlled with the NN-based controller and the gait
equation controller. The video was recorded at 25 frames per
second.
102 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022
50
40
30
20
10
NN-Based
Controller
Conclusion
Designing energy-efficient gaits for snake-like robots remains a
challenging task as they come with redundant DoF and have
complicated interactions with the environment. The most
widely used method for generating gaits for snake-like robots is
the gait equation, which mimics the snake's body shape using
sinusoid-like curves. In this article, we presented a novel gaitdesign
method based on RL. Compared with the optimized
gaits generated by the gait equation, the learned gaits show better
energy efficiencies at medium and high velocities. Although
the learned gaits perform similarly or slightly worse than the
parameterized gaits at low speeds, the proposed method shows
great efficiency in discovering energy-efficient gaits automatically
and maximizing performance by removing the limitations
of predefined models. Our work contributes to and serves as an
exploration for designing sophisticated moving patterns for
snake-like robots. Future work will aim at designing gaits using
RL for snake-like robots without passive wheels.
32
30
28
26
24
22
20
Equation-Based Controller
NN-Based Controller
0.04 0.06 0.08 0.1
Velocity (m/s)
(a)
0.12 0.14
NN-Based Controller
NN-Based Controller (Sim)
0.04 0.06 0.08 0.1
Velocity (m/s)
(b)
0.12 0.14
Figure 8. (a) A real-world power comparison of slithering
gaits from the gait equation and NN-based controllers. The
target velocities are [. ,. ,. ,. ,. ]
velocities are approximately [. ,. ,. ,. ,. ]
0050 10 015020 25 m/s. The actual
0030 05 0080 12 015m/s. The
solid points represent the averaged power over 10 s, and the
error bars depict the standard deviation. Note that the standard
IEEE Robotics & Automation Magazine - December 2022

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