Beruflich Dokumente
Kultur Dokumente
Abstract
Inspired by kangaroos locomotion, we report on developing a kangaroo-style hopping robot. Unlike bipeds, quadrupeds,
or hexapods which alternate the legs for forward locomotion, the kangaroo uses both legs synchronously and generates the
forward locomotion by continuous hopping behavior, and the tail actively balances the unwanted angular momentum generated
by the leg motion. In this work, we generate the Center of Mass (CoM) locomotion of the robot based on the reduced-order
Rolling Spring Loaded Inverted Pendulum (R-SLIP) model, for matching the dynamic behavior of the empirical robot legs. In
order to compensate the possible body pitch variation, the robot is equipped with an active tail for pitch variation compensation,
emulating the balance mechanism of a kangaroo. The robot is empirically built, and various design issues and strategies are
addressed. Finally, the experimental evaluation is executed to validate the performance of the kangaroo-style robot with hopping
locomotion.
Keywords: legged robot, kangaroo, hopping, SLIP, tail
Copyright 2014, Jilin University. Published by Elsevier Limited and Science Press. All rights reserved.
doi: 10.1016/S1672-6529(14)60066-4
1 Introduction
Bio-inspired robotics is a thriving branch of the
robotics family. With emulating from nature in mind,
robotic researchers try to develop or improve the mechanical structure, mechatronic system, control algorithm, and overall locomotion behavior of the robots by
incorporating bio-inspired methodologies. Among all,
legged robotics is a particularly popular field of study;
not only because the legged form is the morphology
adopted by most of ground animals and worth understanding, but also the legged robot has great and obvious
potential to initiate dynamic behavior or to negotiate
rough terrain.
The study of dynamic robotic systems was initiated
by the development of a single-leg hopping robot in the
1980s[1]. Following this, various quadruped and hexapod
robots with dynamic behaviors have been reported. For
example, the quadruped robot Tekken II has
spring-mass-damping systems between the leg joints, as
a mechanism to stimulate viscoelastic characteristics of
muscle tissue[2,3]. The quadruped robot Scout II can be
excited to have dynamically stable bounding gait under
complex situation via simple control laws[4,5]. The
Corresponding author: Pei-Chun Lin
E-mail: peichunlin@ntu.edu.tw
542
543
544
(b)
Hip joint
kt
Fiberglass
Tire tread
(c)
Stance phase
Landing
Flight phase
Liftoff
Landing
545
has more details. With the mass, the leg property and
other geometrical dimensions of the robot, the four intrinsic parameters of the R-SLIP model (r, kt, m, l) can be
computed. Table 1 lists the values of the model parameters for analysis.
Stable running trajectory of the R-SLIP model is
selected based on the return map analysis[31]. As shown
in Fig. 3a, the fixed point exists when the next (i+1)th
touchdown angle (i+1) is the same as the current ith
touchdown angle (i). Moreover, the fixed point is stable
when its slope is between 1 and 1. Otherwise it is considered as an unstable fix point. Fig. 3b plots the distribution of the fixed points within the range if ICs are
achievable by the empirical robot. Ideally, if the robot is
operated near or at the stable fixed points, it moves like
the R-SLIP model, and it can stably run without any
energy input (i.e., purely passive dynamics). However,
in practical the robot is not energy conservative due to
various energy loss terms such as friction and damping,
so the motor power is required to drive the virtual leg
and provides the energy input to the system, so the energy level of the robot can be kept unchanged. Considering stability and empirical motor power rating,
v =1.5 ms1 and = 44 are chosen as the ICs of the
robot. Following this, stance-phase time 0.145 s,
flight-phase time 0.186 s, and lift-off angle 11.52 can be
yielded. The solved mass trajectory of the R-SLIP is
implemented as the ideal CoM trajectory of the robot,
which also equals to the leg trajectory (t). The motion is
highly nonlinear, so it is solved off-line and is approximated by a fifth-order polynomial. The approximated and the original trajectories are plotted in Fig. 4.
The Root Mean Square Error (RMSE) between these
two trajectories is 0.0181, showing that the polynomial
approximation matches the original trajectory quite well.
The leg trajectory during the flight phase can be freely
designed as long as it satisfies the speed and profile
continuity. The algorithm of leg trajectory generation is
also shown as a sub-algorithm of the overall control
system in Fig. 5.
m (kg)
r (m)
kt (Nm)
l (m)
mb (kg)
ml (kg)
mt (kg)
rb (m)
b ()
rl (m)
rth (m)
rtv (m)
5.400
0.100
16.2
0.118
4.606
0.423
0.371
0.017
55.0
0.090
0.120
0.177
546
Fig. 3 Stability analysis of the R-SLIP model: (a) Return map and
(b) distribution of the stable and unstable fixed points. The symbols , , and v indicate the landing angle, touchdown angle, and
touchdown speed, respectively. The filled and hollow legends
represent stable and unstable fixed points, accordingly.
l,set (deg)
60
Theoretical value
Approximated value
40
20
Stance phase
Flight phase
20
0
25
50
Period (%)
(%)
Period
75
100
Fig. 4 The ideal leg trajectory of the robot, shown in dashed red.
The blue curve with triangular markers is the approximated trajectory implemented on the robot. l represents the leg position,
which will be further defined in Fig. 6. The ICs are v = 1.5 ms1,
= 44and = 58. The hopping period is 0.33 s.
its stance phase, ideally the robot moves purely by passive dynamics without motor input, and in that situation
no external torque is generated to alter the body orientation. In contrast, while the robot is in its flight phase,
the motor power should be applied to repose the legs
back to the ICs for the next touchdown. The torque applied to the legs generates unwanted reactive torque
back to the robot and make the body pitch change. In
addition, relative movements of the legs and the tail to
the body also cause the variation of the CoM position.
When the CoM of the overall robot does not locate at the
rotation axis (i.e., hip joint), where the leg is mounted,
the gravity will also generate another unwanted torque to
make the body pitch change. Thus, the motion of the tail
is designed to compensate for the effects caused by these
two factors. This compensation mechanism requires a
dynamic model of the overall robot, including the body,
leg, and tail.
The derivation of the dynamic model is described
as follows. Fig. 6 plots the notations for the model development. While the robot is in its flight phase, the
moment equation can be expressed as
gml rl sin(l + l ) gmb rb cos( + b ) +
(1)
gmt (rth + rtv sin( + )) = I b + I l (l + ) + I t ( + ).
By importing the ICs of these states, the tail trajectory
can be solved numerically. The values of the geometric
parameters for solving Eq. (1) are listed in Table 1. Because Eq. (1) is also highly nonlinear, the trajectory of
the robot tail is solved off-line and is approximated by a
fifth-order polynomial function. The trajectories with
rb
mb, Ib
b
rtv
l
rl
ml, Il
Fig. 7 The ideal tail trajectory of the robot, shown in dashed red.
The blue curve with triangular markers is the approximated trajectory implemented on the robot.
547
548
Two sets of experiments are conducted for performance validation: one is for evaluating the tail performance, and the other is for evaluating the robot hopping motion. For the latter experiment, the motion of the
robot is computed based on the sequential snapshots
taken by a stationary camcorder (HDR-SR11, Sony)
549
RMS error
Stationary tail
2.24
Max range
7.18
1.17
5.27
0.96
4.49
Fig. 10 The kangaroo robot with hopping locomotion. (a) Schematic drawing of the robot with hopping locomotion in one period;
(b) snapshot of the hopping robot.
550
y (m)
y (m)
x (m)
system, such as CoM touchdown velocity, CoM touchdown angle, leg configuration, body pitch, body pitch rate,
etc. Empirical experience on robot experiment runs reveals that it is very hard to provide correct ICs in all states
to the robot by merely throwing the robot to the ground by
an operator or by a sliding guide system. Through
trial-and-error process we found that the feasible solution
of providing roughly adequate ICs to the robot is gradual
transition from vertical hopping to forward hopping. At
the very beginning, the robot is supported by four thin
wires at four corners of the body, which roughly defines
and confines the configuration of the robot at rest. Then
the robot is set to hop. Because of the wires, the robot can
move freely in the vertical direction but constrained in the
forward direction. During the hopping motion the wires
keep at the same configuration, so when the robot lands on
the ground, the body pitch can be regulated to the right
value. In the meantime, the operator also gradually moves
forward to reduce the forward motion constraint to the
robot. After about six transition strides, the robot can
roughly reach the desired ICs. Because of using this state
transition method, the images shown in Fig. 10b has human operator following the robot. It is not perfect but at
least applicable to the hopping experiment.
The quantitative results of the robot with hopping
locomotion are shown in Figs. 11 and 12. Fig. 11 shows
the CoM displacement of the robot versus time in
(a)
20
10
0
10
Robot
R-SLIP model
20
30
(b)
20
60
40
Period (%)
80
100
140
120
100
80
60
40
Robot
R-SLIP model
20
0
0
20
60
40
Period (%)
80
100
Fig. 12 The forward (Fx) and vertical (Fy) ground reaction forces
of the robot (blue curve, with mean (middle blue curves) and
standard deviation (std, top and bottom blue bars)) and the R-SLIP
model (red dashed curves) during stance phase.
551
552
(d)
0.3
0.3
0.2
0.2
y (m)
y (m)
(a)
0.1
0.1
0
0
0.1
(b)
0.2
x (m)
0.3
0.4
0.1
0.2
x (m)
0.1
0.2
(e)
0.3
0.4
20
Fx (N)
Fx (N)
10
0
10
20
12
0.1
(c)
0.2
Time (s)
0.3
30
0.4
0.30.3
0.3
0.4
0.4
0.4
160
140
120
100
80
Fy (N)
Fy (N)
0.4
Time (s)
(f)
140
120
0.3
60
40
20
100
80
60
40
20
0
20
0
20
0
0.1
0.2
Time (s)
0.3
0.4
0 0.1
0
0.1
0.1
0.1
0.2
0.20.20.2
0.3
Time (s)
Fig. 14 The mass trajectory and the ground reaction forces of the R-SLIP model with (ac) varying touchdown angle and = 58 as well
as with (df) varying landing angle and = 44. The touchdown velocity is v = 1.5 ms1.
opposite direction, which makes the body pitch up. Because some discrepancy exists between the robot and the
model (such as leg mass and complex circular leg behavior), the pitch-up phenomenon is observed in the
experiments as shown in Fig. 13. With this undesired
and gradually-increased body pitch, the robot fails to
hop after several strides. As a result, the tentative solution to remedy this accumulated error is to use the
mounted wires to calibrate the body pitch back to the
consistent condition at the lift-off moment. By carefully
checked the sequential images recorded by the camcorder, we confirm that the strings interferes with the
system about 20% time in each stride, which made the
experimental setup not perfect but tolerable. With this
calibration, the robot could hop continuously. Though
several imperfect conditions exist in current methodology, the use of R-SLIP model as the motion template[32]
for the robot still plays an important role of developing
the robot with dynamic behavior. If the robot doesnt
have compliant legs, it is hard to initiate the dynamics.
7 Conclusion
We report on the development of a kangaroo robot
with dynamic locomotion. The use of the R-SLIP model
helps in the morphology design of the robot as well as
initiating its dynamic hopping behavior. The experimental result shows that the robot with the active tail can
significantly reduce the body pitch variation to about
half in comparison to the robot with the stationary tail. In
comparison with the open-loop active tail, the
closed-loop active tail can further improve the performance but not as significant as from stationary tail to
active tail. The second experiment reveals that it is a
challenge to transit the robot from a stationary state to
the designed stable locomotion. Once transient, the robot
can hop for a couple of strides without being tethered.
The accumulated body pitch error is the main cause for
the robot to fail on long time-scale hopping, and this
deviation also induces the tail motion to gradually shift
its operation region from middle to one side of its
achievable motion range. This reveals the facts that the
robot in this configuration may not have wide stable
region, and the carefully-tuned operation point and the
whole-body model may be required for further motion
improvement. Though not perfect, the current design
morphology of the robot allows the robot to induce the
dynamic locomotion of the kangaroo robot, which
clearly exhibits the alternating behavior of stance phase
and flight phase.
We are currently working on revising the mechatronic system of the robot, which can provide more accurate sensory information and stronger computation
553
Acknowledgment
The authors would like to thank Chun-Kai Huang
for discussing the development issues in mathematical
modeling, sensor developing, and programming. This
work is supported by National Science Council (NSC),
Taiwan, under contract NSC 100-2628-E-002-021-MY3.
References
[1]
[2]
[3]
[4]
Poulakakis I, Smith J A, Buehler M. Modeling and experiments of untethered quadrupedal running with a bounding
gait: The Scout II robot. International Journal of Robotics
Research, 2005, 24, 239256.
[5]
[6]
[7]
Cham J G, Karpick J K, Cutkosky M R. Stride period adaptation of a biomimetic running hexapod. International
Journal of Robotics Research, 2004, 23, 141153.
[8]
[9]
[10] Blickhan R. The spring mass model for running and hopping.
554
373378.
kangaroo
778778.
robot
with
dynamic
jogging
locomotion.
en_corp/13704.htm.
429439.
555
4156.