The 2009 IEEE/RSJ International Conference on

Intelligent Robots and Systems

October 11-15, 2009 St. Louis, USA

Real Time Motion Generation and Control for Biped Robot

-3rd Report: Dynamics Error Compensation-
Toru Takenaka, Takashi Matsumoto and Takahide Yoshiike

Abstract—Trajectories generated from approximate dynamics cannot be used in real time.

models can lead biped robots to fall down due to the difference of Kajita et al. [9] proposed a method to compute the translational
dynamics between the approximate dynamics model and the real and rotational velocities of the upper body and free ends to
robot. In this paper, we propose real time methods to compensate achieve desired translational and angular momentum. Kajita et
for the dynamics error using dynamics error compensation
al. [7] also proposed predictive control techniques to
models. Our methods satisfy the horizontal ground reaction force
and moment limits so that no slip is caused with the ground. We compensate for the dynamics approximation error with a
also propose a method to compensate for the knee dynamics error feedback loop. These approaches ignore friction limits of
which is not modeled in our approximate dynamics models. ground and the resulting motion can lead to a slip.
Combining these techniques, running motion is achieved on a real In this paper, we propose methods to compensate for the
biped. dynamics error while explicitly taking the horizontal ground
reaction force limits into account. We use an inverted pendulum
I. INTRODUCTION model and a flywheel model [14] to account for the dynamics
error. Using these techniques, the dynamics error is
compensated and the horizontal ground reaction force limits are
satisfied simultaneously. In addition, we propose a method to
kinematically account for the knee bending motion of the
detailed robot model for the motion generated from an
approximate dynamics model without knees. Using these
techniques, fast motions such as walking and running generated
from approximate dynamics models are realized on a real robot.
In section II, a general overview of the system is given. From
section III to V, the approximation model of the robot dynamics
Fig. 1 Running biped robot system (ASIMO) and gait pattern modification using them are introduced.
Experimental results are shown in section VI.
For biped robots [1][2] (Fig. 1) to exist around and collaborate
with human, they need abilities to react robustly against
II. SYSTEM OVERVIEW
unknown events including avoiding collision with previously
unknown obstacles and maintaining balance under external In this paper, a gait pattern is a set of trajectories for the desired
disturbances by taking steps. Real time techniques to generate a ZMP, the feet and the upper body.
variety of dynamically stable motions are required to achieve Given a command to move, step position and duration are
such behaviors. decided (Fig. 2(a)).
Because of its high degrees of freedom, highly nonlinear Given parameters above, design the desired ZMP and feet
dynamics and kinematic constraints, it is difficult to use trajectories. Then design the upper body trajectory which
detailed dynamics models of biped robots to generate their enables the desired ZMP trajectory without causing the upper
motions in real time. Thus, approximate dynamics models are body to diverge (Fig. 2(b)).
widely used to generate motions in real time [3][4][5][6][7]. Feed the gait pattern into the real robot, and stabilize it while it
Due to the approximated dynamics, motions generated with is following the gait pattern (Fig. 2(c)).
approximate dynamics model cannot be followed exactly by the Walking and running gait generations are explained in [10] and
real robot. Yamane et al. [8] and Nagasaka [13] have proposed [11] respectively, and the balance controller is explained in [12].
methods to minimize the dynamics error using optimization This paper addresses the dynamics error compensation (Fig.
techniques. These approach require large computation time and 2(e)) in detail. The dynamics error compensation consists of the
knee dynamics compensation (Fig. 2(f)) and the dynamics error
Fundamental Research Center, Honda Research & Development, 8-1 Honcho, compensation with the ground reaction force limits (Fig. 2(g)).
Wako-shi, Saitama, Japan.

978-1-4244-3804-4/09/$25.00 ©2009 IEEE 1594

 Gait pattern generator (b) h : height of the pendulum.
Foot step planner Motion generation Instantaneous model
(a) by simple model error compensation
Actual robot
g : gravitational acceleration constant.
(d) (e) (c)
stability
Even though the pendulum accelerates vertically, we assume
control the vertical motion is small enough and can be approximated by
a constant. As implied in Eq. (3), the inverted pendulum
requires the ground reaction force to generate the ground
Knee dynamics
compensation
Full model
compensation
(e)
reaction moment.
(f) (g) The flywheel model (Fig. 3(b)) represents the motion of the
robot in which moment is generated while the CoG stays still.
Iθ&&wheel = M wheel (4)
Fig. 2 System overview where I is the characterized from the real robot.
On the real robot, inclining the upper body forward without
III. GAIT PATTERN MODIFICATION WITH DYNAMICS ERROR translating it causes the CoG to move forward. To mimic the
COMPENSATION MODELS behavior of the flywheel, the upper body of the robot has to be
We design the desired ZMP trajectories with approximate shifted horizontally by x wheel as follows so that the CoG of the
dynamics models. If a detailed dynamics model tries to follow robot does not move.
these trajectories, horizontal ground reaction moments are x wheel = Cθ wheel (5)
generated about the desired ZMP due to different dynamics. We where C is a constant identified from an upright posture of the
measure this dynamics error as the horizontal ground reaction real robot.
moment generated by the detailed dynamics model about the
desired ZMP trajectory designed with the approximate models. Horizontal Model (a) Rotational Model (b)

This moment is zero if a trajectory generated from and followed

m pend I
by the same model, and thus we use it as a measure of dynamics
error between two different dynamics models.
Instead of treating the dynamics difference directly, we use h
models representing the position and the angle of the upper
body to cancel the horizontal ground reaction moment to 0. One
of the advantages of this approach is that the kinematic x ZMP
pend
constraints and redundancy of the detailed dynamics models
can be ignored. Another advantage is computational efficiency
which is critical to satisfy real time requirement. Redundancy of
the legs of the detailed dynamics model is not discussed in this
paper.
The dynamics error compensation models are an inverted
pendulum and a flywheel (Fig. 3). In this paper, we limit our
(a) (b)
discussions to motions in saggital plane, but similar
compensation is done in the lateral plane as well on a real robot. Fig. 3 Dynamics error compensation models
The motion of the inverted pendulum model (Fig. 3(a)) is
governed by the following dynamics equations.
IV. FEEDFORWARD AND FEEDBACK DYNAMICS
m pend h&x&pend = m pend ( g + &z&pend ) * x pend + M pend (1)
COMPENSATION METHODS
Fpend = m pend &x&pend (2)
∆M pend = h∆F pend (3) In this section, the following three feedforward (FF) methods
where and a feedback (FB) method to compensate for the dynamics
x pend : horizontal position of the pendulum. error between approximate dynamics models and the real robot
are explained.
m pend : mass of the pendulum. 1. FF compensation method which ignores ground reaction
Fpend : ground reaction force of the pendulum. force limits
2. FF compensation method which satisfies ground reaction
M pend : ground reaction moment of the pendulum.
force limits.
&z&pend : vertical acceleration of the pendulum. 3. FF compensation method which satisfies both ground
reaction force limits and moment limits.

1595
Only the third FF method is implemented in our system, but the Approximate
Detailed dynamics
model + xbody _ mdfd
other two FF methods are used for the ease of explaining. dynamics model xbody _ cmd F full + Fmdfd
Inverted +
A. Feedforward Dynamics Error Compensation without θ body _ cmd M full pendulum model
x pend
Ground Reaction Force Limits - Fpend
Pendulum Stability +
Fig. 4 is the block diagram of the first compensation method control
M pend _ stab M mdfd
which ignores the ground reaction force limits. This is
equivalent to Kajita et al. [7] except that we do not use preview Fig. 4 Feedforward dynamics error compensation
control. without ground reaction force limits
The desired ZMP trajectory x zmp , feet trajectory and upper body
xbody _ mdfd
position xbody _ cmd and inclination θ body _ cmd are designed using an +
+
xbody _ cmd F full Fmdfd
approximate dynamics model. These are transformed into +
θ body _ cmd M full x pend
desired angle for each joint through inverse kinematics on the
- Fpend
detailed dynamics model. Now, the moment about the desired
M pend _ stab Pendulum Stability +
M
ZMP, M full ,and the ground reaction force, F full can be control
ltd
pend _ stab z −1
Z −1 Fbias +
-
computed. M full is caused due to the difference between the
Fpend _ stab + unltd
Ftotal Ftotal
ltd z −1 M mdfd
1 -
two dynamics models and thus ideally is 0. The goal of the h + +
h
( = M ltd
pend _ stab )
dynamics compensation is to make M full to 0 using the inverted
Fig. 5 Feedforward dynamics error compensation with
pendulum model introduced earlier. This can be achieved by ground reaction force limits in discrete time
substituting − M full into M pend in Eq. (1) and adding resulting
x pend into xbody _ cmd . This usually causes the inverted pendulum B. Feedforward Dynamics Error Compensation with
Ground Reaction Force Limits
to diverge. To prevent it, a stabilizing control which outputs
Fig. 5 is the block diagram of the second compensation method,
M pend _ stab is enhanced to make x pend 0. The stabilizing
an extension from the first method to satisfy the ground reaction
controller can be a simple PD controller for example. An force limits.
example controller for running motion is introduced in [12]. According to Eq. (1)(2), if M ltd is increased by
pend _ stab ( = M mdfd )
The sum of requested dynamics compensation moment and
∆M ltd , then Fmdfd increases by ∆M ltd pend _ stab h
.
pendulum stabilizing moment, (− M full + M pend _ stab ) , is then pend _ stab

Fmdfd = M pend _ stab / h + Fbias

ltd
(6)
input to the pendulum, and the output x pend is added to xbody _ cmd
where Fbias is the instantaneous value of Fmdfd when M pend _ stab ltd

to obtain the modified desired upper body position xbody _ mdfd .

becomes zero. Assuming that the value of Fbias did not change
M pend _ stab prevents the inverted pendulum from converging, between the previous and current time steps, the expected value
however it also becomes the moment about the original desired unltd
Ftotal of Fmdfd without limit in the current time step can be
trajectory. Thus, M pend _ stab becomes the ground reaction
obtained as follows.
moment of modified gait pattern M mdfd . The control scheme unltd
Ftotal = M pend _ stab / h + Z −1 Fbias (7)
shown in Fig. 4 tries to make the difference of M full and M mdfd where Z −1 Fbias is the value of Fbias from the previous time step.

to 0. M mdfd behaves like low-pass filtered M full so it fluctuates

unltd
Ftotal is input to the limiter in Fig. 5 which outputs a limited
value, Ftotal
ltd
, between [Fmin , Fmax ] . Using the output from the
less.
The modified gait pattern generates desired ground reaction limiter, the pendulum stabilizing moment M ltd pend _ stab
is
force, Fmdfd , approximately equal to F full + F pend . Even if F full determined.
−1
pend _ stab = h ( Ftotal − Z Fbias )
M ltd ltd
is designed so that it satisfies the limits, there is no guarantee (8)
that the modified force Fmdfd does not exceed the limits.

1596
 + xbody _ mdfd From Eq. (9)(10)(12),
+
xbody _ cmd F full Fmdfd M mdfd = M ltd
pend _ stab + M wheel _ stab
ltd (13)
θ body _ cmd M full x pend
+
Note that Fmdfd is obtained from Eq. (6).
Fpend
- In case the ground reaction force and horizontal moment do not
Pendulum
M pend _ stab
Stability
+
exceed the limits, the expected values of Fmdfd and M mdfd ,
control Z −1 Fbias +
-
unltd
( Ftotal unltd
, M total ) , can be computed as shown in Fig. 9. These
Fpend _ stab + Ftotal
unltd
Ftotal
ltd M mdfd
1 - KF
h values are input to the limiter which modifies them to Ftotal
ltd
and
h + + s Fpend
ltd
_ stab
M total so that they do not violate admissible values. The details
ltd
Fig. 6 Feedforward dynamics error compensation with
ground reaction force limits in continuous time of how the limiter works are explained in the next section.
Note that M ltd
pend _ stab
is computed in the same way as in the
+ xbody _ mdfd
xbody _ cmd F full +
Fmdfd second method and M wheel
ltd
_ stab
is computed as follows.
+ +
θ body _ cmd M full x pend
ltd
M wheel _ stab = M total − M pend _ stab
ltd ltd (14)
- Fpend The modified gait pattern has admissible values of Fmdfd and
+
M mdfd by designing M ltd
pend _ stab
and M wheel
ltd
_ stab
as explained
above.
- KF M mdfd
Fmax h + + xbody _ mdfd
+ s
+ + + θ
Fig. 7 Feedforward dynamics error compensation with xbody _ cmd F full +
body _ mdfd

+ Fmdfd
ground reaction force limits for Ftotal
unltd
exceeding Fmax +
θ body _ cmd M full
x pend
Note that M ltd
pend _ stab
is the desired ground reaction moment of M ltd - Fpend
pend _ stab

the compensated gait pattern. +

+
M pend M mdfd
Limiter
Fig. 6 is the result of transforming Fig. 5 into continuous time & Distributor M ltd
+
xwheel
wheel _ stab
domain. In case Ftotal
unltd
exceeds Fmax , Fig. 6 transforms into Fig. M wheel θ wheel
7 which ensures that Fmdfd follows Fmax with integral feedback. M pend _ stab
Pendulum Stability
control
M wheel _ stab
C. Feedforward Dynamics Error Compensation with Wheel Stability
control
Ground Reaction Force and Horizontal Moment Limits
Fig. 8 Feedforward dynamics error compensation with
Fig. 8 is the block diagram of the third compensation method ground reaction force and moment limits
which extends the second method to satisfy horizontal moment
limits. Fmdfd
z −1
The limiter and distributor (shaded box in Fig. 8) takes Z −1 Fbias +
-
M pend _ stab , M wheel _ stab and Fmdfd as inputs and outputs M ltd
pend _ stab M pend _ stab 1 + Ftotal
unltd
Ftotal
ltd
- z −1 M ltd
pend _ stab
moment h
and M wheel
ltd
_ stab
. Using the outputs, the pendulum and the h + +
ltd
M wheel _ stab + M unltd force
M total
ltd - M wheel _ stab
flywheel models are given following inputs. total +
+
M pend = − M full + M ltd (9)
pend _ stab Fig. 9 Force and moment limiter and distributor
M wheel = M wheel
ltd
_ stab (10) in in
Moment Moment ( M total , Ftotal )
As shown in Fig. 8, the modified upper body motion which the lmtd lmtd
( M total , Ftotal )
real robot tries to follow is a sum of the trajectory designed with
the approximate dynamics model, the pendulum motion and the M max M max
flywheel motion. As a result, the desired ground reaction force Fmin Fmax Force Fmin Fmax Force
is also a sum of the forces from the approximate dynamics M min M min
model and the two compensation models.
Fmdfd = F full + F pend (11) (a) (b)
M mdfd = M full + M pend + M wheel (12) Fig. 10 Force and moment modification by the limiter

1597
 D. Ground Reaction Force and Momentum Limiter dynamics model and the detailed dynamics model using an
The rectangle in Fig. 10 is the admissible range of the ground inverted pendulum and a flywheel as dynamics compensation
reaction force and moment. The limiter outputs the input models. However, the integrated dynamics of the approximate
(Ftotal
unltd unltd
, M total ) without modification if its coordinate is inside dynamics model, inverted pendulum model and flywheel model
do not strictly match that of the detailed dynamics model.
this rectangle. On the other hand, if the point (Ftotal
unltd unltd
, M total ) lies Fig. 11 is the block diagram for the FB method. The motions of
outside the triangle, M wheel _ stab is modified in such a way that
ltd
dynamics error models, an inverted pendulum and a flywheel,
ltd
M wheel becomes as close as possible to M wheel _ stab . Inputting are added to the original gait pattern and input to the detailed
_ stab
dynamics model. On the other hand, the motions of the
moments which differs from M wheel _ stab by large amount causes
dynamics error models in the FF methods are not input to the
the flywheel to diverge from its original state thus leading the detailed dynamics models. The desired ground reaction
upper body of the robot to incline largely. As shown in Fig. moment of the modified gait pattern, M mdfd , is set to
10(a), a line which goes through (Ftotal ) and has (M _ stab )
unltd unltd
, M total .
ltd
pend _ stab + M wheel
ltd

inclination h is used to make the modification. Modification is

M err , the difference of M full and (M ltd
pend _ stab + M wheel _ stab )
ltd
is
made in such a way that Eq. (3) holds. As long as Eq. (3) is true,
the motion can be realized by the inverted pendulum. Otherwise, integrated and input to an inverted pendulum model which has
the flywheel has to rotate to generate a moment which causes the same physical properties as the model used in the FF
ltd
M wheel to diverge from M wheel _ stab . methods.
_ stab
Compared to the FF method, the FB method is robust against
In case the line goes through the rectangle, the intersection
the modeling error of the dynamics compensation models, the
between the edge of the rectangle and the line which is closer to
inverted pendulum and flywheel. However, the FF method has
the original coordinate becomes the output. If the line does not
no delay in compensating the dynamics error while the FB
intersect with the rectangle, the point on the edge of the
method does.
rectangle which is closest to the line becomes the output.
From these properties of FF and FB methods, we serialize the
The admissible values of Fmdfd are determined based on the two systems to achieve small delay and high accuracy. First the
friction between the ground and the sole of the feet and the FF method eliminates large portion of the dynamics error
admissible values of M mdfd are determined so that the ZMP without delay, and then the FB method prevents the remaining
error from accumulating.
stays inside support polygon. Note that these admissible values
are dependent on the vertical motion of the robot. During the
V. KNEE BENDING DYNAMICS COMPENSATION
flight phases when the vertical ground reaction force is 0, both
lmtd
M total and Ftotal
lmtd
are 0. The approximate dynamics models we use to generate the
original gait pattern do not have components corresponding to
the knee bending motion of walking or running. This can cause
xbody _ mdfd
θ body _ mdfd
large dynamics error especially for motions in which the knees
xbody _ cmd +
F full
of the robot bends quickly. Thus, we compensate for the knee
θ body _ cmd
+
+
M full motion using the following method and then feed the
+
+
-
M err compensated motion into the FF and FB methods explained
KM
s
previously.
M lmtd - M pend x pend Fig. 12 shows the models we use to compensate for the knee
pend _ stab

+
+ +
+ M mdfd bending dynamics. Fig. 12(a) is used to generate the original
Limiter
& Distributor lmtd
+
xwheel
gait pattern, Fig. 12(c) is the detailed dynamics model of our
M wheel M wheel
_ stab
θ wheel real robot and the Fig. 12(b) is the telescopic model. The
M pend _ stab
Pendulum Stability control
telescopic model has thigh, shin and upper body links. Each of
M wheel _ stab the links has the same mass and inertia as those of the
Wheel Stabilitycontrol
corresponding link of the detailed dynamics model. The CoG’s
of the thigh and shin links are located on the line which
Fig. 11 Feedback dynamics error compensation
connects the hip joint and ankle joint.
Note that the GoG of the upper body link of the gait pattern
E. Feedback Dynamics Compensation
designed with the approximate dynamics model is x s b and the
The previous section explained methods to modify gait patterns
angle from vertical line is θ s b . Also, the CoG of the thigh, shin
to compensate for the dynamics error between an approximate
and upper body links of the detailed dynamics model are

1598
 θ cb
located at x c t [i ] , x c s[i ] and x c b ([i] is either [R] or [L] for left θ tb

and right leg respectively). The angles of each link are θ c t [i ] , hcb hc b

θ c s[i ] and θ c b . The masses are mt , ms and mb , and the inertias

θ ct
hc t hc t
are I t , I s and I b . For the telescopic model, the CoG of the θ tt
θ cs
thigh, shin and upper body links of the telescopic model are hc s hc s
θ ts
located at x t t [ i ] , x t s[i ] and xt b ([i] is either [R] or [L] for left and
right leg respectively). The angles of each link are θ t t [i ] , θ t s[i ] ZMP

and θ t b . The vertical CoG of each link are hc t , hc s and h c b Fig. 13 Translational and rotational knee dynamics
respectively and assumed to be constant during motion. These compensation
values are determined from a standard upright posture of the
robot in which the knees are bent at some angle. The position of VI. RESULTS
the ankle joint of the detailed dynamics model and telescopic We conducted experiments using a biped robot who is 130 cm
model are identical to that of the approximate dynamics model. tall. Each leg of the robot has 6 degrees of freedom. Fig. 14 is
Note also that xt b and θ t b are identical to x s b and θ s b the moment generated about the desired ZMP trajectory
respectively. transformed into equivalent ZMP error while walking at 4 km/h.
We solve the following two kinematic constraints iteratively for As seen, the motion generated from the approximate dynamics
x c b and θ c b to compensate for the knee dynamics. x c b and θ c b error without compensation generates maximum error of 110
are the modified upper body position and angle respectively. mm(A). With the knee dynamics compensation, the maximum
error decreases to 90 mm(B). Furthermore, with the FF method
mb ( x c b − xt b ) + mt ∑ x c t [i ] − xt t [i ] ( ) following the knee dynamics error compensation and the FB
i
method, the maximum error is less than 20 mm. The result
(
+ ms ∑ x c s[i ] − xt s[i ] = Const. ) (15)
confirms that proposed methods successfully compensates for
i
mb h c b × ( x c b − x t b ) dynamics error of the approximate dynamics model and realizes
(
+ mt ∑ h c t [ i ] × ( x c t [ i ] − x t t [i ] ) ) motion with large margin of stability.
Fig. 15 shows the behavior of the inverted pendulum and
i

+ m ∑ (h s
c
s[ i ] × ( x c s[ i ] − x ) ) + I (θ − θ )
t
s[ i ] b
c
b
t
b
flywheel dynamics error compensation models with and
i without knee dynamics compensation while running at 6km/h.
+ I ∑ (θ
t
c
t[i ] − θ t t[i ] ) + I ∑ (θ − θ ) = Const.
s
c
s[ i ]
t
s[ i ] (16) Without the knee dynamics compensation, both dynamics error
i i compensation models diverge. With the knee dynamics, no
Regarding the difference of horizontal positions of each mass as divergence is observed. The result indicates effectiveness of the
velocity, Eq. (15) means that the total moment of the masses proposed knee dynamics error compensation technique. It also
stays constant (Fig. 13(a)(b)). Eq. (16) refers to constant total indicates that there is a limit on how much dynamics error the
angular momentum (Fig. 13(c)). From these kinematic feedforward and feedback compensation techniques can
computations, the dynamics error due to the knee bending compensate for as explained in previous sections.
motion is decreased. 120
(A)Dynamics model with three mass
(a) (b) (c)
θ t
b θ c
b 80
Ib mb (C)FF type
mb
xcb
xt b 40
ZMP error [mm]

θ t t [[ii ] It θ c t [i ]
mt
x t t [i ] x c t [i ] 0

ms
Is
x t s[[ii ] x c s[[ii ] -40
θ t s[[ii ] θ c s[i ] (D)FB type
(B)Knee bend
-80
Fig. 12 Telescopic model and knee dynamics model
-120

Double support phase (80) Single support phase (320) Time [ms

Fig. 14 ZMP error of 4km/h walking

1599
 60
Another method to compensate for the motion of knees which
our approximate dynamics model did not have is introduced. It
Pendulum position [mm]
FB type without knee bend FB type with knee bend

was then shown that the knee dynamics compensation is an

0
effective way to prevent the dynamics error compensation
FF type without knee bend
models of the feedforward and feedback approaches to diverge.
FF type with knee bend
-60
Time
20
REFERENCES
FF type without knee bend
[1] K. Hirai, M. Hirose, Y. Haikawa, and T. Takenaka, “The Development of
Wheel position [deg]

FB type without knee bend

Honda Humanoid Robot”, In Proceedings of the 1998 IEEE International
0 Conference on Robotics and Automation，Leuven，Belgium，May.
1998，pp. 1321-1326.
FB type with knee bend
FF type with knee bend [2] Y. Sakagami, R. Watanabe, C. Aoyama, S. Matsunaga, N. Higaki, and K.
Fujimura, “The intelligent ASIMO: System overview and integration”, In
-20
Time Proceedings of the 2002 IEEE/RSJ International Conference on
Fig. 15 Dynamics error model compensation for Intelligent Robots and Systems, 2002，pp. 2478-2483.
[3] S. Kajita, O. Matsumoto, and M. Saigo, “Real-time 3D Walking Pattern
running at 6km/h Generation for a Biped Robot with Telescopic Legs”, In Proceedings of
15 the 2001 IEEE International Conference on Robotics and Automation,
Knee bend 2001, pp. 2299-2306.
FF type [4] T. Sugihara, Y. Nakamura, and H. Inoue, “Realtime Humanoid Motion
Generation through ZMP Manipulation based on Inverted Pendulum
Moment [Nm]

0 Control”, In Proceedings of the 2002 IEEE International Conference on

Robotics and Automation, pp. 1404-1409, Washington DC, May, 2002.
FB type [5] K. Nagasaka, Y. Kuroki, S. Suzuki, Y. Itoh, and J. Yamaguchi,
“Integrated Motion Control for Walking, Jumping and Running on a
-15 Small Bipedal Entertainment Robot”, In Proceedings of the 2004 IEEE
Time International Conference on Robotics and Automation, New Orleans,
300
2004, pp. 3189-3194.
[6] K. Harada, S. Kajita, K. Kaneko, and H. Hirukawa, “An Analytical
reaction force [N]

Method on Real-time Gait Planning for a Humanoid Robot”, Journal of

Horizontal ground

Humanoid Robotics, vol.3, no.1, pp.1-19, 2006.

0
[7] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi, and
H. Hirukawa, “Biped Walking Pattern Generation by using Preview
Control of Zero-Moment Point”, In Proceedings of the 2003 IEEE
International Conference on Robotics and Automation, 2003, pp.
-300
Time 1620-1626.
[8] K. Yamane and Y. Nakamura, “Dynamics Filter - Concept and
Fig. 16 Moment and horizontal ground reaction
Implementation of On-Line Motion Generator for Human Figures”, IEEE
force of running at 6km/h Transactions on Robotics and Automation, vol. 19, no. 3, pp. 421-432,
Fig. 16 shows the dynamics moment error and designed 2003.
horizontal ground reaction force for running at 6 km/h [9] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi, and
H. Hirukawa, “Resolved Momentum Control: Humanoid Motion
expressed in the same unit as in Fig. 15. The three lines show Planning based on the Linear and Angular Momentum”, In Proceedings
dynamics error after the knee bending compensation, another of the 2003 IEEE/RSJ International Conference on Intelligent Robots and
followed by FF method, and another followed by FF and then Systems, Oct. 2003, pp. 1644-1650.
[10] T. Takenaka, T. Matsumoto, and T. Yoshiike, “Real Time Motion
FB methods. It can be observed that the ground reaction Generation and Control for Biped Robot -1st Report: Walking Gait
moment about the desired ZMP has decreased by the proposed Pattern Generation-”, In Proceedings of IEEE/RSJ International
compensation methods. Note that the ground reaction moment Conference on Intelligent Robots and Systems, 2009
[11] T. Takenaka, T. Matsumoto, T. Yoshiike, and S. Shirokura, “Real Time
error is 0 (Fig. 16, top) and the designed horizontal ground Motion Generation and Control for Biped Robot -2nd Report: Running
reaction force (Fig. 16, bottom) is also 0 during the flight phase. Gait Pattern Generation-”, In Proceedings of IEEE/RSJ International
Conference on Intelligent Robots and Systems, 2009.
VII. CONCLUSIONS [12] T. Takenaka, T. Matsumoto, T. Yoshiike, T. Hasegawa, S. Shirokura, H.
Kaneko, and A. Orita, “Real Time Motion Generation and Control for
In this paper, we proposed methods for correcting dynamics Biped Robot -4th Report: Integrated Balance Control-”, In Proceedings of
error of trajectories generated from approximate dynamics IEEE/RSJ International Conference on Intelligent Robots and Systems,
2009.
model of the real robot. The feedforward and feedback [13] K. Nagasaka, “The Whole Body Motion Generation of Humanoid Robot
approaches using an inverted pendulum and a flywheel to Using Dynamics Filter (in Japanese) ”, Ph.D. dissertation, University of
approximate the dynamics error were proposed. The two Tokyo, 2000
[14] J.Pratte, J.Carff, S.Drakunov and Ambarish Goswami, “Capture Point: A
methods keep the horizontal ground reaction force and moment Step toward Humanoid Push Recovery”, In Proceedings of the 2006
under limits to prevent a slip between the foot and the ground. IEEE International Conference on Humanoids，2006，pp. 200-207.

1600