Вы находитесь на странице: 1из 8

Electrical Engineering in Japan, Vol. 169, No.

2, 2009
Translated from Denki Gakkai Ronbunshi, Vol. 127-D, No. 8, August 2007, pp. 884889

Trajectory Planning and Control of One-Legged Hopping Robot at Thrusting Phase

SHO SAKAINO and KOUHEI OHNISHI


Keio University, Japan

SUMMARY switching from one of these functions to another. By allo-


cating an individual actuator to each of these functions they
One of the most important events for hopping robots facilitated analysis of the control process [2]. Based on the
is lift-off (the end of thrusting phase) since hopping motions method proposed by Raiberts group, several scientists
often become unstable during the lift-off. Therefore, we succeeded in implementation of hopping robots [3, 4]. The
focus on the thrusting phase. In this research, we propose a disadvantage of the Raibert method, despite simplicity of
new controller, which combines two controllers. We also application of control rules, is that it can be used only with
propose a trajectory planning so that a hopping robot fol- robots equipped with direct action hydraulic actuators
lows command value of the vertical velocity. The effective- which compromises control robustness and makes it diffi-
ness of the proposed methods is confirmed by simulation
cult to accomplish accurate hopping motions. We think that
and experiments. 2009 Wiley Periodicals, Inc. Electr Eng
accuracy of hopping motions can be improved by using a
Jpn, 169(2): 2936, 2009; Published online in Wiley Inter-
robot having legs with rotary joints driven by motors char-
Science (www.interscience.wiley.com). DOI 10.1002/
acterized by a better controllability. Morita and Ohnishi
eej.20831
proposed a method of application of Raiberts method to
robots having rotary joints and achieved continuous hop-
Key words: hopping robot; jumping robot; hop-
ping action [5]. However, the reality is that compared to
ping height control; thrust force.
robots with direct action joints, robots with rotary joints
have a more serious problem with model uncertainties due
1. Introduction to widely varying position of the center of gravity. There-
fore, it was impossible to achieve highly accurate operation
Legged humanoid robots can move over rough ter- of the hopping robots with rotary joints even when the
rains with uneven surfaces, go over obstacles, climb stairs, MoritaOhnishi method was used.
and make sharp turns. These properties make them suitable In this study, we first designed, using the Raibert
for operation in the human environment. Patterns of human method, a simple-model control system based on an as-
locomotion include walking and running. So far there have sumption that the robot leg is massless. Then, we proposed
been several reports on the designing of walking robots and
a method of adding a feedback input to compensate for
the next goal is to teach humanoid robots to run. Running
uncertainty of modeling the position of the center of gravity
is a moving action when both feet momentarily leave the
(COG). Compensation of the error of the COG position
ground (hopping). A specific characteristic of running is
according to this method makes it possible to apply the
the ability to cover a distance greater than permitted by the
structural limitations of the robot. Therefore, a running Raibert method to the robot having rotary joints and to
robot can move faster and to greater distances than a walk- accurately control the hopping height. Another important
ing robot. factor of the stable hopping is the trajectory planning. In the
First studies on jumping robots are attributed to Mat- conventional method, command values are applied discon-
suoka [1]. Raibert and colleagues using direct action hy- tinuously and the switching of controllers causes instability
draulic actuators developed a robot having a structure based of the robot. Therefore, for the purpose of a smooth con-
on a concept of the massless leg. They proposed a new troller switching in this study it was decided to introduce a
method that consisted in dividing the control of the hopping trajectory planning for the purpose of following commands
process into three independent functions (hopping height of the reference lift-off velocity. A possibility of imple-
control, body attitude control, and speed control) and in menting the accurate hopping height control by means of

2009 Wiley Periodicals, Inc.


29
the proposed control system and the trajectory planning are
demonstrated by simulation and experiments.
In this paper, we first discuss a model of the hopping
robot and explain existing and proposed control methods.
Results of the study are illustrated by simulation and ex-
periments.

2. Model of the Hopping Robot


Figure 1 shows the hopping robot used in this study.
Model of the world coordinate frame is shown in Fig. 2; l1
through l6 denote link lengths, m1 through m6 denote masses Fig. 1. One-legged robot.
of the links; the coordinate origin is set at the lower end of
link 1. This robot has two rotary joints: 2 and 3. The robot
is connected to the counterweight m3 by the shaft of link 3
and can rotate around the passive joint 1. The purpose of
the counterweight is to offset the force of gravity. Due to
this shaft, the robot operation is restricted to the sagittal
plane. Model of the sagittal plane is shown in Fig. 3.
Coordinates of the sagittal plane are defined as the base
coordinate system and its origin is set in joint 2.
The hopping cycle consists of three main events:
landing, lowest point, and lift-off. Periods between
these events are divided into three phases defined as flight
phase (from lift-off to landing), touchdown phase
(from landing to the lowest point), and thrusting
phase (from the lowest point to lift-off). By switching Fig. 2. World coordinates.
the controller from one phase to another, it is possible to
produce complicated hopping motions by a simple control
system [3]. In this paper, we propose to introduce a thrust-
ing phase controller and the trajectory planning. Definitions
of scripts used in this paper are given in Table 1.

2.1 Kinematics of the hopping robot

(1)

Position of the leg tip bxtip in the base coordinates of the


hopping robot is expressed by Eq. (1). From Eq. (1), one Fig. 3. Sagittal plane.
can derive Eq. (2) giving the relationship between velocity
of the leg tip and the angular velocity of the joint and Eq.
(4) giving the relationship between velocity of the leg tip
and the angular acceleration of the joint. Table 1. Scripts

(2)
(3)
(4)

30
2.2 Dynamics of the hopping robot Equation (6) is derived based. on geometrical relationships
.
as a transform from wzg to 3. From this relational expres-
In this paper, the motion equation of the hopping sion, a controller D.for 3 can be structured by deriving the
.
robot in the two-dimensional space is given by Eq. (5) [6, term wzcmd
g from wcmd
3 . This is discussed in the following
7]: section.
(5)
4. Proposed Method

In this paper we proposed a control system combin-


ing the world coordinate controller and the base coordinate
controller as the control system of the vertical velocity to
improve control of the thrusting phase. We also proposed a
trajectory planning for a better compliance of the lift-off
velocity with its reference value.

3. Conventional Method
4.1 Compensation of model uncertainties
According to the Raibert method, during the thrust-
The cross mark in Fig. 5 shows the COG position of
ing phase the hip joint 2 controls the attitude and the knee
. an actual robot. Since the model shown in Fig. 4 was
joint 3 controls the vertical velocity z. Since in robots with
assumed to be massless, the COG position was assigned to
rotary joints responses of all axes generally interfere with
joint 2, but since legs of a real robot have mass, the COG
each other, the stance control and the vertical velocity
position was shifted to a location lower than joint 2. There-
control also interfere with each other. However, this inter-
fore, lleg becomes shorter than the actual leg. The problem
ference can be avoided by using a disturbance observer [8].
was that the derived value of the lift-off velocity was lower
The goal of this study is to improve robustness of the
than the desired velocity.
vertical velocity control. According to the MoritaOhnishi
In conventional methods, control in the base coordi-
method, if the hopping robot shown in Fig. 4 is considered
nates is carried out .by the feedback control by calculating
a point mass, then the vertical velocity can be controlled by
the response value res3 of the joint angle rather than using
command values of the knee joint [5]. Let us assume that
the reference command value wzcmd in the world coordi-
the center of gravity is in the point shown in this figure by
nates. However, model uncertainties appearing in the world
the cross mark, then by denoting variables in this diagram
coordinate frame cannot be compensated in the base coor-
as
dinates. On the other hand, it is possible to use the feedback
control in world coordinates by directly calculating the
response values wzres in the world coordinates. Compared
to the control in the base coordinates, control in the world
we have coordinates is more robust against model uncertainties of
the COG position. However, since the absolute location of
(6) humanoid robots is usually determined by image process-
ing or PSD (Photoshop document), observation errors are

Fig. 4. Robot model in thrusting phase. Fig. 5. COG position.

31
greater than when control is carried out in the base coordi-
nates because the accuracy of the computer processing is
lower than that of the encoder that plays the role of the
sensor in the base coordinates.
As indicated above, control in base coordinates is
characterized by large model uncertainties requiring a high-
accuracy sensor, while the world coordinate control is
characterized by absence of model uncertainties but it is
usually accompanied by large observation errors. In other
words, these two types of control are in a trade-off relation-
ship. In this study we propose a method combining these
two control methods. Fig. 6. Block diagram.
First, the D control regarding 3 is given by the
formula

(7) right side of Eq. (9). It is understood that two control


. systems are synthesized based on the selection ratio .
.
Here, considering that res
3 is obtained from the value of Figure 7 shows a simplified block diagram of the

w res
.3 obtained in the world coordinates and the value of proposed method. C denotes Eq. (10) and the robot is
3 obtained in the base coordinates, Eq. (7) can be
b res denoted as G. It is assumed that the world coordinate
transformed based on the selection ratio as follows: control is free of the above model uncertainties and only
observation errors are present, while observation errors in
the base coordinates control are relatively small and only
(8) modeling uncertainties are present. (1 )D is control in
the base coordinates and N is the errors occurring in the
world coordinate control. It is also assumed that distur-
bances taking place in the robot, such as friction, are sup-
On the other hand, by using Eq. (6), the following relation-
ship can be derived: pressed by the disturbance observer and their effect is
disregarded in this study. Therefore, the transfer function
from command values wzcmd g to response values wzresg is
(9) expressed by

(12)
Control in the world coordinates is carried out using this
equation. Since Eq. (9) does not comprise the term lleg,
control in the world coordinates is not affected by model Transfer functions of (1 )D and N to respective re-
uncertainties of lleg associated with the modeling of the sponse values can be expressed by the equations
massless leg. By substituting Eq. (9) in Eq. (8), we obtain
(13)

(14)

(10)

(11)

thus combining control in base coordinates and world co-


ordinates. Value of this selection ratio is determined
experimentally. Block diagram of this control system is
shown in Fig. 6. In the diagram, A is the denominator of the Fig. 7. Simplified block diagram.

32
Effect of errors depends on the selection ratio . Therefore, function having four unknown quantities. If the time when
the effect of errors on the response in the proposed control- the hopping robot is in its lowest position t = 0 and the time
ler can be expressed by when the robot achieves the desired velocity t = Tth, then
the velocity command values of the cubic function satisfy-
(15) ing the above conditions are applied in the interval of 0 t
Tth. If the robot did not lift off before the lift-off reference
time, then the velocity command in the interval Tth < t is to
One of the problems of designing the controllers is that it
maintain the velocity constant.
is usually necessary to compromise between the robustness
If the initial velocity and the initial acceleration at t
toward model uncertainties and the robustness toward ob- . ..
= 0 are given by wzres 0 and wzres0 , respectively, then the
servation errors [9]. Design of an optimal controller be-
trajectory expressed by the cubic function satisfying the
comes even more difficult when nonlinear coordinate
above-mentioned conditions can be expressed by
transform, like in Eq. (6), is involved. However, according
to Eqs. (12) and (15), in the proposed controller, transfer
function C from command values to response values and
transfer function from errors to response values can be
designed independently as two transfer functions by adjust-
ing the ratio. This means that the controller and the
trade-off arrangements can be designed separately and
since the trade-off is determined by , the system can be
optimized by easy adjustments. According to this method, (16)
an optimal trade-off can be searched even in the case of
linear control. Since it is difficult to formalize D and N, it
is impossible to determine the numerical value of . There- Figure 8 shows velocity command as a function of time. Tth
fore, the value of was found experimentally. In this study value is determined experimentally. The reference value of
..
was a constant but the selection ratio can also have band the lift-off acceleration wzcmd is 0 and the reference lift-off
. lo

characteristics similar to LPF. velocity wzcmdlo is determined using the OhashiOhnishi


method [10] by the expression

4.2 Trajectory planning (17)

In this section we propose the trajectory planning for .


Here, wzpretop is the COG velocity of the previous step at the
the vertical velocity. The desired hopping height is achieved lift-off time, wzpre
top is the hopping height of the previous step,
by tracking the reference lift-off velocity of the robot. In and wztop is the desired hopping height. When the robot
the conventional method proposed by Morita and Ohnishi, moves along this trajectory, no large impact is produced at
during robot operation, discontinuous commands are ap- the time of the control switching, thus making it possible
plied at the time of switching from the touchdown phase to to implement stable hopping.
the thrusting phase. Consequently, at the initial moment of
the thrusting phase, a strong impulse force is applied to the
robot, which may result in the loss of the robot stability. The
purpose of the proposed trajectory planning is to maintain
continuity of the trajectory during the period from the
touchdown phase to the thrusting phase. Since, as follows
from the motion equation, the acceleration can be obtained
by dividing the force by the mass, there is a close relation-
ship between the force and the acceleration. Therefore, it is
necessary to suppress the impulse force accompanying the
strong feedback by applying the command values concate-
nated to response values up to the acceleration dimension.
Conditions satisfying the proposed velocity trajec-
tory are: the matching of velocity and acceleration to the
response values as the initial condition and the matching of
velocity and acceleration to their reference values as the
final condition. Since it is necessary to satisfy these four
conditions, velocity command values are given by a cubic Fig. 8. Velocity command.

33
5. Simulation Table 2. Robot parameters

5.1 Touchdown control

In this study, the soft landing trajectory was adopted


as the trajectory for the touchdown phase [11]. The control
system used during the touchdown phase was the variable
compliance control based on the soft landing trajectory
employed by Ohashi and Ohnishi [12].

5.2 Floor model


Table 3. Control parameters
Floor model used in this study was a springdamper
model which makes it possible to accurately set parameters.
The counteracting force Fext added at the contact point can
be denoted by

(18)

Table 4. Environment parameters

During the period when the legs tip is in contact with the
floor surface (zf 0 Fext,z 0) the value of xf0 is main-
tained unchanged.

5.3 Results of simulation

Lengths and masses of robot links are listed in Table


2. Other parameters used in simulation are given in Tables
3 and 4. Gg is the cutoff frequency of the disturbance
observer, St is the sampling time. To reproduce the obser- Fig. 9. Velocity response of conventional method.
vation error of the vertical position, the resolution of wzres
g
was set to 0.3 mm.
The proposed controller for the thrusting phase
shows an improved velocity compliance. Figure 9 shows
velocity response according to the conventional method and
Fig. 10 shows the velocity response of the proposed con-
troller. These graphs depict response during the thrusting
phase. The COG velocity is plotted along the vertical axis
and time is plotted along the horizontal axis. The left end
of the graph corresponds to the time when the COG is at its
lowest position and the right end corresponds to the lift-off
phase. In both cases the robot lifts off 0.13 s after being in
the lowest COG position. Dotted lines correspond to re-
sponse values and solid lines show command values. Robot Fig. 10. Velocity control of the proposed method.

34
Table 5. Specifications for motors

Fig. 11. Vertical COG position in simulation.

trajectory in both cases is given by Eq. (16). Because of the


effect of model uncertainties, the deviation between the
velocity response and the command is almost 7 m/s in the
conventional method, but when the proposed controller is
used, the deviation was confirmed to amount to only 5 m/s,
thus considerably improving the tracking of velocity com-
mands.
Figure 11 depicts results of simulation of the re-
sponse of the vertical COG position with the proposed
controller. Time is plotted along the horizontal axis. The Fig. 12. Vertical COG position in experiment ( = 0.0).
highest COG position specified by the command is 0.27 m.
Trajectory planning according to the conventional Morita
Ohnishi method can easily result in a loss of the robots
stability, thus making it difficult to maintain continuous
hopping. The proposed method was confirmed to be able to
continuously sustain the hopping of the robot.

6. Experiments

Experiments were carried out by measuring 1 with


the encoder accuracy reduced to 1/10 of the nominal value,
thus achieving the resolution of the absolute position wz of
approximately 0.3 mm, which is almost equal to that of Fig. 13. Vertical COG position in experiment ( = 0.5).
PSD. Control parameters other than are the same as in
Table 3 and specifications of motors used in the experimen-
tal robot are given in Table 5. Experimental results at
different values of are shown in Figs. 12, 13, and 14. The
reference hopping height is 0.38 m and the trajectory during
the thrusting phase is given by Eq. (16).
At = 0.0, which corresponds to the conventional
method, the hopping height oscillates around the reference
value. At = 1.0, the hopping height became erratic and
after the tenth hop the robot fell down. At = 0.5, oscilla-
tions characteristic of the robot with conventional control
disappear and the deviation from the reference value of the
hopping height becomes smaller than in the conventional
method. These results confirm that the proposed trajectory
planning makes it possible to accomplish steady hopping. Fig. 14. Vertical COG position in experiment ( = 1.0).

35
7. Conclusion Conference on Robotics and Automation, Leuven,
Belgium, p 793798.
In this paper we proposed a control system combin- 5. Morita Y, Ohnishi K. Attitude control of hopping
ing the world coordinates controller and the base coordi- robot using angular momentum. Proceedings of the
nates controller. This makes it possible to compensate for IEEE International Conference on Industrial Elec-
model errors of the thrust period and to achieve a better tronics, Maribor, Slovenia, p 173178, 2003.
following of the vertical velocity, thus implementing a 6. Paul RP. Robot manipulator. MIT Press; 1981.
highly accurate hopping height. In addition, we proposed a 7. Brady M et al. Robot motion. MIT Press; 1982.
continuous trajectory planning of the vertical velocity up to 8. Ohnishi K. Robust motion control by disturbance
the acceleration dimension. Effectiveness of these two pro- observer. J Robotics Soc Japan 1993;11:486493. (in
posed methods was confirmed by simulation and experi- Japanese)
ments. 9. Shimizu K, Ohmori H. Introduction to theory of
linear control. Baifukan Press; 2003. (in Japanese)
REFERENCES 10. Ohashi E, Ohnishi K. A hopping height control for
hopping robot. IEEJ Trans Ind Appl 2004;124:660
1. Matsuoka K. Fundamental study on hopping and 665. (in Japanese)
running mechanisms. Trans Japan Soc Mech Eng 11. Nakano E, Okubo H, Kimura H. The landing control
1977;43:45014509. (in Japanese) of a jumping machine. J Robotics Soc Japan
2. Raibert MH. Legged robots that balance. MIT Press; 1991;9:169176. (in Japanese)
1986. 12. Ohashi E, Ohnishi K. Variable compliance control
3. Ahmadi M, Buehler M. Stable control of a simulated based on soft-landing trajectory for hopping robot.
one-legged running robot with hip and leg compli- Proceedings of the 30th Annual Conference of the
ance. IEEE Trans Robotics Automation 1997;13:96 IEEE International Electronics Society, Busan, Ko-
104. rea, 2004.
4. Zeglin G, Brown B. Control of a bow leg hopping
robot. Proceedings of the 1998 IEEE International

AUTHORS (from left to right)

Sho Sakaino (student member) graduated from the Department of System Design Engineering, Keio University, in 2006
and is now involved in research on hopping robots.

Kouhei Ohnishi (senior member) received his Ph.D. degree from the Department of Electrical Engineering, University
of Tokyo, in 1980 and joined the Department of Electrical Engineering, Keio University, as a research associate. He is now a
professor in the Department of System Design Engineering.

36

Вам также может понравиться