Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids)

July 19-21, 2021. Munich, Germany

Whole-body walking pattern using pelvis-rotation for long stride


and arm swing for yaw angular momentum compensation
Beomyeong Park1 Myeong-Ju Kim 1 Eunho Sung1 Junhyung Kim1 and Jaeheung Park2

Abstract— A long stride can enable a humanoid robot achieve to the movement of the swinging leg or foot. The pelvis
fast and stable walking. For a long stride, the kinematics of rotation trajectory is generated in the form of a sine function
the robot should be fully utilized, and walking with pelvic so that the hip of the swinging leg can follow the swinging
rotation can be a solution. A rotational trajectory of pelvis
considering kinematic limitations is needed for pelvis-rotation foot [5]. In addition, the rotation angle of the pelvis was
walking. When the robot walks with a long stride while rotating generated proportional to the distance between the support
the pelvis, the yaw momentum may be larger than that when foot and swing foot, or in proportion to the velocity of
walks with the pelvis fixed. This is caused by the rotation of the the swing foot [6], [7]. These methods did not reflect the
pelvis and leg motion, and hence, walking with pelvic rotation kinematic constraint of the robot, but it was possible to
may become unstable. In this paper, we propose to control the
lower body of a robot as a redundant system with leg joints increase the stride by heuristically generating the rotational
and a waist joint for walking with pelvic rotation. The position trajectory of the pelvis.
of the base frame to implement the redundant system for the The optimization technique is used to generate the pelvis
lower body of the robot is also proposed. In addition, the a rotation trajectory while satisfying the constraint and min-
quadratic programming (QP) controller is formulated to enable imizing the objective function. A method was introduced
arm swing for yaw momentum compensation while controlling
the lower body. The feasibility of the proposed control method to obtain the rotational trajectory of pelvis that minimizes
was verified using a simulation and an experiment of walking the yaw momentum at the support foot while reducing the
with a long stride while rotating the pelvis using a QP controller error in the trajectory of the pelvis obtained from motion
and compensating the yaw momentum by means of arm swing. capture data [8]. In the study, a rotational pelvic trajectory
I. INTRODUCTION obtained using motion capture was required. A method of
obtaining the three-dimensional pose trajectory of pelvis
The humanoid robot can walk stably, but the slow walk- using a genetic algorithm was introduced [9]. In this method,
ing speed degrades the efficiency of the robot to perform the rotation of the pelvis was not calculated directly through
tasks [1]. Therefore, in order to increase the practicality of optimization, but by using the optimized pelvis and foot
bipedal walking, the necessity of studies for fast walking is positions.
inevitable. To walk fast, the robot should be able to walk with For the robot to walk fast, yaw momentum caused by
a long stride. Heel-strike and toe-off walking was studied to swing leg movement must be compensated. If the yaw
increase the stride of the robot by overcoming the kinematic momentum is greater than the rotational friction, the robot
limitations [2], [3]. However, when a flat foot is used that is will slip. To compensate the yaw momentum of the robot, the
not designed for heel-strike and toe-off walking, the walking momentum compensation using the waist joint to rotate the
may become unstable as the edge of foot contact the ground. torso was studied [10]–[12]. In biomechanics, the effect of
Pelvis-rotation walking can be a way for a robot with a arm swing on the angular momentum in human gait was
flat foot to increase stride length without the risk of contact analyzed in [13], [14]. In order to apply the momentum
instability in heel-toe walking. reduction due to arm swing to the robot, a simplified model
Analysis of human walking patterns using motion capture of the robot was used to generate the arm motion. Using the
revealed that human walking involves pelvic rotation [4], [5]. Simplified upper body model, the trajectory of rotating the
In humanoid robot research, walking with pelvic rotation has torso or swinging the arm for compensating the momentum
been studied to increase the stride and walk like a human. was generated [15]. However, since the momentum of the
The trajectory of the pelvic rotation was generated according robot was not calculated in detail, the yaw momentum was
This work was supported by the National Research Foundation of not perfectly compensated. To overcome the limitation of
Korea (NRF) grant funded by the Korea government (MSIP) (No. the simplified model, a centroidal momentum that projects
NRF-2015R1A2A1A10055798), and the Korea government(MSIT) (No. the momentum of each link generated by its joint velocity
2021R1A2C3005914) and the Technology Innovation Program (10060081)
funded by the Ministry of Trade, industry & Energy (MI, Korea). ※ MSIT to the COM coordinate of the robot was introduced [16].
: Ministry of Science and ICT Centroidal momentum was used not only for whole body
1 Beomyeong Park, Myeong-Ju Kim, Eunho Sung, and Junhyung Kim
controller for balancing control, but also for generating arm
are with Graduate School of Convergence Science and Technology,
Seoul National University (on2lord,myeong-ju,Eunho526 swing motion to control the momentum while walking [17]–
and john3.16)@snu.ac.kr [19]. To calculate the centroidal momentum, the momentum
2 Jaeheung Park is with Department of Transdisciplinary Studies, Seoul
of each link is accumulated and added from each end-effector
National University, 08826 Seoul, Republic of Korea, and Advanced In-
stitutes of Convergence Technology, Suwon, 16229, Republic of Korea to the base frame according to the robot tree structure. The
park73@snu.ac.kr momentum generated by the waist yaw joint in the base

978-1-7281-9371-7/21/$31.00 ©2021 IEEE 47


𝜃𝜃𝑝𝑝𝑝𝑝𝑝𝑝
𝜃𝜃𝑝𝑝𝑝𝑝𝑝𝑝
left left
footfoot 𝑙𝑙𝑝𝑝𝑝𝑝𝑝𝑝𝑙𝑙𝑝𝑝𝑝𝑝𝑝𝑝
𝑤𝑤𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
𝑤𝑤𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
y yx x pelvis
pelvis right footfoot
right
𝑙𝑙𝑝𝑝𝑝𝑝𝑝𝑝 sinsin
𝑙𝑙𝑝𝑝𝑝𝑝𝑝𝑝 𝜃𝜃𝑝𝑝𝑝𝑝𝑝𝑝
𝜃𝜃𝑝𝑝𝑝𝑝𝑝𝑝 ∆𝑥𝑥∆𝑥𝑥
𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓
𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓
z z

z z
x x
y y

𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 ∗ ∗
𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
𝑙𝑙𝑚𝑚𝑚𝑚𝑚𝑚
𝑙𝑙𝑚𝑚𝑚𝑚𝑚𝑚 𝑙𝑙𝑚𝑚𝑚𝑚𝑚𝑚
𝑙𝑙𝑚𝑚𝑚𝑚𝑚𝑚

(a) Pelvis-fixed walking (b) Pelvis-rotation walking


Fig. 2: Kinematic chain tree with suggested base frame. The
Fig. 1: Kinematic analysis of pelvis rotation. Each figure joints robot are divided two parts : lower body and upper
shows x − y view on the top and x − z view on the bottom. body.

frame located on the pelvis represents the momentum of pelvis-fixed can be calculated as follows.
the torso. The yaw momentum calculated from the base step
q
lmax 2 − z2
= 2 lleg (1)
frame located in pelvis does not reflect the momentum that 0
occurs due to the rotation of the pelvis link in pelvis-rotation When the pelvis rotates θpel around the z axis, the distance
walking. ∆Xf oot which the swing foot can move further is determined
In this paper, we propose a method that enables pelvis- in proportion to the width of pelvis, lpel and rotation angle
rotation walking by using the redundant characteristic that of pelvis, θpel .
occurs when controlling the lower body with 13 joints. The
location of the base frame was suggested to make the lower ∆Xf oot = lpel · sinθpel (2)
body of the robot into a redundant system. The proposed base
When pelvis rotates as much as it coincides with the
frame allows the momentum generated by pelvic rotation to
straight line connecting the ankles of the two feet, the maxi-
be reflected when calculating centroidal momentum. Further,
mum kinematic stride of pelvis-rotation walking is achieved.
an arm swing motion is generated to compensate for the q
yaw momentum The whole-body walking motion including step∗ step 2
lmax = (lmax + lpel )2 − wstep (3)
pelvis-rotation walking and arm swing is generated using the
QP controller. The Dyros-Jet humanoid robot used in this study has lleg
The remainder of this paper is organized as follows. In of 0.7742 m, z0 of 0.6881 m, lpel of 0.21 m, and wstep
Section II, we analyze the stride that can be increased when of 0.243 m, respectively [20]. In the pelvis-fixed walking,
step step∗
rotating the pelvis, and proposes a position of the base frame. lmax is 0.5669 m, and in pelvis-rotation walking, lmax is
Section III introduces the system architecture. In section 0.7269 m, a stride increases by 22% compared to pelvis-fixed
IV, the objective functions and constraints that constitute walking.
the proposed quadratic programming (QP) controller are de- B. Redundant system of lower body with base frame on waist
scribed. In Section V, simulation and experimental results are
Most humanoid robots use the base frame on the pelvis
discussed, and in Section VI, the conclusions are provided.
coordinate. In this base frame, the waist yaw joint rotates the
II. KINEMATIC ANALYSIS torso. However, in this paper, the waist yaw joint is used to
rotate the pelvis. A base frame is set on the waist yaw link to
In this section, the merit of pelvis rotation walking is ana- use 12 leg joints, qL,L and qR,L and 1 waist yaw joint qW Y
lyzed kinematically. The tree structure of a robot including a for pelvis-rotating walking control. Fig. 2 shows the position
proposed position of base frame for pelvis-rotation walking of the proposed base frame and the relationship between
is described. joints. Based on the proposed base frame, the lower body
joint is composed of 13 degrees of freedom (DOF), qlow ∈
A. Kinematic advantage of pelvis-rotation walking  T
qW Y qL,L qR,L , and the upper body is composed of
The stride that increases due to pelvis rotation can be ana- 15 joints
 including 1 waist T roll and 7 DOF for each arm,
lyzed kinematically. Fig. 1 compares and analyzes the pelvis- qup ∈ qW R qLA qRA .
fixed walking and pelvis-rotation walking. To calculate the
maximum kinematic stride of the robot, it is assumed that III. SYSTEM STRUCTURE OVERVIEW
the pelvis is in the center of the two feet. If the length of This section describes the proposed system structure. The
the leg is lleg and the height from ankle to hip is z0 when stride and distance are used as input of the walking pattern
step
the knee is straight, the maximum stride length, lmax , when generator. The footstep planner generates foot holds. The

48
IV. QUADRATIC PROGRAMMING CONTROLLER
FOR PELVIS ROTATING AND ARM SWINGING
A QP controller for pelvis-rotation walking using a redun-
dant lower body and for compensating yaw momentum using
arm swing will be described in this section. The output of
QP controller is q̇, and objective functions and constraints
that formulate the QP controller are introduced.
A. Multi-Objective Function
The optimization variable of the QP controller, q̇ ∈ Rn+13 ,
is composed of lower body joints, q̇low ∈ R13 and selected
upper body joints q̇sel,up ∈ Rn . The selected joint, q̇sel,up
is selected from upper body joints q̇up ∈ R15 using the
selection matrix Ssel,up ∈ Rn×15 .
 T
q̇ = q̇low q̇sel,up (7)
T
Fig. 3: System structure of proposed method.

q̇low = q̇W Y q̇L,L q̇R,L (8)
q̇sel,up = Ssel,up · q̇up (9)

trajectory of COM is generated by using MPC to minimize , where value of Ssel,up(i,j) corresponding to the selected j
the COM velocity fluctuation [21]. The position trajectory of joint is 1 and the rest are 0.
the base frame is generated to track the trajectory of COM, 1) Desired trajectory tracking: The trajectory tracking
and the orientation trajectory of the base frame is generated objective function allows the robot to be controlled according
to maintain the same orientation as that of the foot hold. to the desired foot trajectory described in Section III.
Foot trajectory is generated as a third-order polynomial so w1
f1 (q̇) = kJlow · q̇low − ẋblow,d k2 , (10)
that the robot foot can move to a predetermined foot hold. 2
b
Foot velocity trajectory basedi on the base frame, Ẋf,d = where
h  
b
ẋbf,d ẏf,d b
żf,d φ̇ θ̇ ψ̇ , is calculated from the foot JLf,1 JLf,2−7 0
Jlow =
trajectory and the base trajectory. JRf,1 0 JRf,2−7
b
T
ẋlow,d = ẋLf,d ẋbRf,d
 b
QP controller receives foot Jacobian Jf , angular momen-
tum L, and measured joint angle q̇ from robot. Each foot
is controlled by 7 joints by the tree structure introduced in The Jlow ∈ R12×13 is lower body Jacobian matrix, Jf,1 ∈
6×1
Section II-B, and the Jf is as follows. R is waist yaw joint Jacobian martix, and Jf,2−7 ∈ R6×6
is the Jacobian matrix of each leg, respectively. ẋblow,d ∈ R12
is desired foot velocity vector.
" b b b b b b b
#
∂Pf ∂Pf ∂Pf ∂Pf ∂Pf ∂Pf ∂Pf
Jf = ∂q0 ∂q1 ∂q2 ∂q3 ∂q4 ∂q5 ∂q6 , (4) 2) Regularization: If the optimization variable, q̇, changes
b b b b b b b
Z0 Z1 Z2 Z3 Z4 Z5 Z6 suddenly, the robot may become unstable. Therefore, in order
to prevent rapid changes in q̇, an objective function that
where Pbf ∈ R3 is position vector from base to each foot, serves as a regularization is used.
q0 is the waist yaw joint and q1 to q6 are the joints ordering w2
from hip to ankle of each leg. The b Zj ∈ R3 are the rotation f2 (q̇) = kq̇k2 (11)
2
axis of joint j expressed in the base frame.
3) Recovering initial pose: The swinging arm motion is
The relationship with q̇ is expressed using the Jacobian of
generated by the momentum constraint will be explained in
each foot and the desired foot velocity.
next section. However, due to the kinematic asymmetry of
the robot or the uncertainty of control, the movement to
ẋbLf,d = JLf · q̇L,low , (5)
compensate for the momentum may drift in either direction.
ẋbRf,d = JRf · q̇R,low , , (6) The objective function to prevent drift is as follows,
w3
 T  T f3 (q̇) = kqd,upper − qinit,upper k2 , (12)
where q̇L,low = q̇W Y q̇L,L , q̇R,low = q̇W Y q̇R,L 2
are left and right leg joint velocity vector, respectively.  
where qupper = qW Y qup is including waist yaw joint
The QP controller, which will be introduced in the next and selected upper arm joints, qd,upper and qinit,upper are
section, outputs the joint velocity q̇, and q̇ is integrated to desired and initial joint angle of qupper , respectively.
become q. The compensator calculates ∆q for compensating
the elasticity of the actuator and for compliant contact [22], B. Constraint
[23]. The robot is controlled by the desired joint angle, q∗ , The QP controller for pelvis-rotation walking include task,
obtained by adding ∆q to q. momentum, and hardware constraints that must be satisfied.

49
(a) Pelvis-fixed walking
(a) Pelvis-rotation walking without arm swing

(b) Pelvis-rotation walking with arm swing


(b) Pelvis-rotation walking with arm swing
Fig. 4: Simulation snapshots of pelvis-fixed and pelvis-
rotation walking. Fig. 5: Experimental snapshots of pelvis-rotation walking
with proposed QP controller

1) Task Constraint: Two leg restraints for walking include


the support foot fixed to the ground and the swing foot C. QP controller
following the swing foot trajectory. Therefore, the constraints The QP controller is configured using the multiple objec-
of the support foot and swing foot are defined as follows. tive functions and constraints introduced in section IV-A and
Jsup · q̇sup = ẋbsup , (13) IV-B. QP controller uses n+13 size of optimization variable
q̇ of (7). Equation (10) and (12) are reorganized with q̇ and
Jsw · q̇sw = ẋbd,sw , (14)
included in the QP controller.
where Jsup and Jsw are Jacobian matrix of support leg and w1
swing leg, and ẋbsup and ẋbsw are desired velocity of support f1 (q̇) = kJ · Ssel,low · q̇ − ẋd k2 , (18)
2
foot and swing foot, respectively. w3
f3 (q̇) = kSsel,upper · (qd − qinit )k2
2) Yaw Momentum Constraint: If the yaw momentum 2
w3
caused by the leg motion is completely compensated by the = kSsel,upper · q̇ · ∆t + Ssel,upper (qcur − qinit )k2 ,
swing of the arm, the yaw momentum caused by the robot’s 2
(19)
joint becomes zero. Therefore, the following constraints are
used for yaw momentum compensation. where Ssel,low ∈ R13×(n+13) is selection matrix for lower
body joints, Ssel,upper is selection matrix for waist yaw joint
ACM M,z · q̇ = 0, (15)
and selected arm joints.
1×(n+3)
where ACM M,z ∈ R is the yaw component of the QP controller for pelvis-rotation is defined as follows
centroidal momentum matrix, and ACM M · q̇ = Lz is a using objective function and constraint defined above.
yaw angular momentum, Lz expressed on COM coordinate
k
[18]. With the proposed position of the base frame, the waist X
min fi (q̇)
yaw joint rotates the pelvis link, so the calculated angular q̇ i=1
momentum reflects the momentum caused by the rotation of
the pelvis. s.t. Task Constraint (13)(14),
(20)
3) Kinematic Constraint: Kinematic constraint specifies Momentum Constraint (15),
the limit of the velocity of each joint and limits the range Joint angle limit (16),
motion of the joint angle. For example, in order to prevent Joint velocity limit (17),
the knee from bending in the opposite direction, the knee
angle is limited beyond the straight leg. where k is the number of objective function. QP controller
in (20) outputs the q̇ which satisfies constraints while mini-
qmin − qcur q − qcur mizing multiple objective functions.
≤ q̇ ≤ max , (16)
∆t ∆t V. DISCUSSION OF SIMULATION AND
q̇min ≤ q̇ ≤ q̇max , (17) EXPERIMENT
where qmin and qmax are minimum and maximum joint The proposed method was validated through simulation
angle limit respectively, q̇min and q̇max are minimum and and experiment. A Dyros-Jet robot with 12 leg joints, two
maximum joint velocity respectively. waist joints, 14 arm joints, and two neck joints was used.

50
(a) The waist yaw joint angle of pelvis-rotation walking.
Fig. 7: Comparison of yaw momentum without or with
arm swinging. The yaw momentum is expressed at COM
coordinate.

(b) The left shoulder pitch joint angle of pelvis-rotation


walking with arm swinging.

(a) Measured z axis torque of on left foot.

(c) The right shoulder pitch joint angle of pelvis-rotation


walking with arm swinging.
Fig. 6: The desired and measured joint angle of pelvis-
rotation walking in experiment.
(b) Measured z axis torque of right foot.
Fig. 8: Comparison of z axis torque of pelvis-rotation walk-
Equation (20) was solved using the QP solver QPOASES ing with or without arm swing.
[24]. To complete the computation of the QP problem at
a rate of 200 Hz, which is the control frequency of the
robot, shoulder joint of each arm were selected, n = 2, so fixed. Therefore, the TORO was able to walk with long
q̇ ∈ R15 is used as optimization variable. In simulations and strides, while avoiding singularity. If the proposed method
experiments, the height from the ground to the base of the includes toe-off motion and the height of the COM is not
robot was 0.8 m. The values of w1 , w2 , and w3 in (20), are fixed, it is expected that a longer stride will be possible.
1, 0.5, and 0.5, respectively. In the experiment, the yaw momentum compensation
In the simulation, the stride lengths of pelvis-fixed walking resulting form the arm swing was analyzed while verifying
and that of pelvis-rotation walking were compared in Fig. 4. pelvis-rotation walking. With or without swinging the arm,
The robot started walking with a 10 cm stride per 1.2 s, and the robot could walk with a stride of 30 cm while rotating
the stride increased by 5 cm. The robot walked a distance of the pelvis. Fig. 6 shows the angle of the waist joint, left
2 m. In the case of walking with a fixed pelvis orientation, a and right shoulder joints. This angle of the waist joint varied
singularity problem arose when the stride was greater than 25 from 2◦ to 8◦ depending on the stride. Fig. 7 shows the yaw
cm. However, with the proposed method, the robot walked angular momentum at COM when pelvis-rotation walking.
with a 40 cm stride with pelvis-rotation. Without the arm swing, the yaw momentum ranged from
Fig. 5 depicts the experimental verification of the proposed 0.34 kg · m2 /s to a maximum of 1.6 kg · m2 /s owing
method. To verify the proposed method experimentally, the to leg movement as the swing stride increased. However,
robot started walking with a stride of 10 cm and then when the robot walks with pelvic rotation and arm swing,
increased the stride to 30 cm and walked a total of 2 m. yaw momentum was compensated and a maximum yaw
This seems to be a shorter stride compared to the 60 cm momentum was decreased to 0.4 kg · m2 /s. In the pelvis-
stride recently demonstrated by the TORO using Heel-toe rotation walking, the RMS of the yaw momentum was 0.5168
[25]. When the COM moved forward, the TORO walked without the arm swing, but with arm swing, and RMS
while dragging his hind leg with the toe of his hind foot decreased by 74.4 % to 0.1324.
touching the ground. Also, the height of the COM was not Fig. 8 shows the z-axis torque, τz measured by the force

51
torque (F/T) sensor attached between the foot and the ankle, [3] M. A. Hopkins, D. W. Hong, and A. Leonessa, “Compliant locomotion
and depicts the change in torque with or without arm swing. using whole-body control and divergent component of motion track-
ing,” in IEEE Int. Conf. Robot. Autom.(ICRA), 2015, pp. 5726–5733.
The relationship between τz at the support foot and Lz at [4] S. Zhang, Q. Huang, H. Wang, W. Xu, and Z. Yu, “The mechanism
COM is as follows [26]. of yaw torque compensation in the human and motion design for
humanoid robots,” Int. J. Adv. Robot. Syst., vol. 10, no. 1, p. 57, 2013.
τz = M (x − Xzmp )ÿ − M (y − Yzmp )ẍ + L̇z , (21) [5] W. Zhang, Q. Huang, Z. Yu, G. Huang et al., “Human-like walking
patterns with pelvic rotation for a humanoid robot,” in Proceeding of
where M is total mass of robot, x and y are position of COM, World Congress on Int. Control Autom. IEEE, 2014, pp. 1887–1892.
and Xzmp and Yzmp are position of ZMP, respectively. [6] C. Zhou, X. Wang, Z. Li, D. Caldwell, and N. Tsagarakis, “Exploiting
the redundancy for humanoid robots to dynamically step over a large
In Fig. 8, the τz increased as the stride increased regardless obstacle,” in Int. Conf. Int. Robot. Syst. IEEE, 2015, pp. 1599–1604.
of arm swing. When the stride was 30 cm without arm swing, [7] J. Lim, I. Lee, I. Shim, H. Jung, H. M. Joe, H. Bae, O. Sim, J. Oh,
after 8 s, the maximum absolute value of τz measured at each T. Jung, S. Shin et al., “Robot system of drc-hubo+ and control
strategy of team kaist in darpa robotics challenge finals,” J. Field
foot was 19.5 N m for the left foot and 19.15 N m for the Robot., vol. 34, no. 4, pp. 802–829, 2017.
right foot. On the other hand, when the stride was 30 cm [8] H. Takemura, A. Matsuyama, J. Ueda, Y. Matsumoto, H. Mizoguchi,
with arm swing, the maximum absolute value of τz of each and T. Ogasawarahi, “Momentum compensation for the dynamic walk
of humanoids based on the optimal pelvic rotation,” in Climbing and
foot was 16.2 N m for the left foot and 14.4 N m for the Walking Robots. Springer, 2006, pp. 485–492.
right foot. The maximum absolute value of τz of each foot [9] Z. Huang, Z. Wang, J. Wei, J. Yu et al., “Three-dimensional posture
decreased by 17 % and 24.8 %, respectively. optimization for biped robot stepping over large ditch based on a
ducted-fan propulsion system.”
As shown in Fig. 7, as Lz at COM decreases remarkably, [10] B. Ugurlu, J. A. Saglia, N. G. Tsagarakis, and et al, “Yaw moment
L̇z also decreases. Since L̇z in (21) was reduced by arm compensation for bipedal robots via intrinsic angular momentum
swing, the τz at single support (SSP) was also reduced. constraint,” Int. J. Humanoid Robot., vol. 9, no. 04, p. 1250033, 2012.
[11] H.-o. Lim, S.-h. Hyon, S. A. Setiawan, and A. Takanishi, “Quasi-
However, the reason that the τz did not decrease significantly human biped walking,” Robotica, vol. 24, no. 2, p. 257, 2006.
seems to be that the torque generated by the position and [12] J. Yamaguchi, E. Soga, S. Inoue, and A. Takanishi, “Development of
acceleration of COM and ZMP in (21) is greater than a bipedal humanoid robot-control method of whole body cooperative
dynamic biped walking,” in Proceedings Int. Conf. Robot. Autom.,
the torque generated by L̇z . Therefore, for stable and fast vol. 1. IEEE, 1999, pp. 368–374.
walking, the trajectories of COM and ZMP considering τz [13] H. Herr and M. Popovic, “Angular momentum in human walking,”
is necessary. Journal of experimental biology, vol. 211, no. 4, pp. 467–481, 2008.
[14] S. H. Collins, P. G. Adamczyk, and A. D. Kuo, “Dynamic arm
VI. CONCLUSIONS swinging in human walking,” Proceedings of the Royal Society B:
Biological Sciences, vol. 276, no. 1673, pp. 3679–3688, 2009.
This paper proposes a whole-body walking pattern gen- [15] J. Ueda, K. Shirase, Y. Matsumoto, S. Oda, and T. Ogasawa, “Mo-
eration method involving pelvis-rotation walking for a long mentum compensation for fast dynamic walking of humanoids based
on pelvic rotation of contact sport athletes,” in Int. Conf. Humanoid
stride and proposes arm swing motion generation for yaw Robots, 2004., vol. 2. IEEE, 2004, pp. 592–607.
momentum compensation. The stride length increased for [16] D. E. Orin and A. Goswami, “Centroidal momentum matrix of a
pelvis-rotation walking compared that for pelvis-fixed walk- humanoid robot: Structure and properties,” in Int. Conf. Int. Robot.
Syst. IEEE, 2008, pp. 653–659.
ing; the scenarios was analyzed kinematically. For realizing [17] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi,
pelvic rotation, the lower body of the robot was controlled and H. Hirukawa, “Resolved momentum control: Humanoid motion
as a redundant system, including the waist yaw joint and leg planning based on the linear and angular momentum,” in Proceedings
Int. Conf. Intell. Robot. Syst., vol. 2. IEEE, 2003, pp. 1644–1650.
joints. For the lower body of the robot to have redundancy, [18] D. E. Orin, A. Goswami, and S.-H. Lee, “Centroidal dynamics of a
the base frame was set to the waist yaw link instead of humanoid robot,” Autom. robots, vol. 35, no. 2-3, pp. 161–176, 2013.
the pelvis. The yaw momentum of the robot was calculated [19] A. Miyata, S. Miyahara, and D. N. Nenchev, “Walking with arm
swinging and pelvis rotation generated with the relative angular
based on the proposed base frame. A QP controller was used acceleration,” IEEE Robot. Autom. Letters, vol. 5, no. 1, pp. 151–158.
to control the redundant lower body to generate the pelvis- [20] S. Park, J. Sim, and J. Park, “System design of humanoid robot dyros-
rotation walking and to generate the arm swing motion for jet,” in IEEE/SICE Int. Symp. System Integration, 2019, pp. 746–750.
[21] B. Park and J. Park, “Walking pattern generation using mpc with
yaw momentum compensation. Simulations and experiments minimization of com velocity fluctuation,” in IEEE Int. Conf. Control,
verified that pelvis-rotation walking with arm swing using Autom. and Syst.(ICCAS), 2020, pp. 268–273.
the proposed method can increase the stride length and [22] M. Kim, J. H. Kim, S. Kim, J. Sim, and J. Park, “Disturbance observer
based linear feedback controller for compliant motion of humanoid
compensate the yaw momentum. The arm swing significantly robot,” in Int. Conf. Robot. Autom. IEEE, 2018, pp. 403–410.
reduced the yaw angular momentum at COM, but the z-axis [23] J. Kim, M. Kim, and J. Park, “Improvement of humanoid walking
torque at the support foot was not significantly reduced. As control by compensating actuator elasticity,” in 2016 IEEE-RAS 16th
Int. Conf. Humanoid Robots, pp. 29–34.
a future study, we will study the walking pattern that not [24] H. Ferreau, C. Kirches, A. Potschka, H. Bock, and M. Diehl,
only compensates for angular momentum of COM, but also “qpOASES: A parametric active-set algorithm for quadratic program-
considers minimization of torque at the support foot. ming,” Math. Program. Comput., vol. 6, no. 4, pp. 327–363, 2014.
[25] G. Mesesan, J. Englsberger, G. Garofalo, C. Ott, and A. Albu-Schäffer,
R EFERENCES “Dynamic walking on compliant and uneven terrain using dcm and
passivity-based whole-body control,” in Int. Conf. Humanoid Robots.
[1] S. Kim, M. Kim, J. Lee, and et al, “Team snu’s control strategies for IEEE, 2019, pp. 25–32.
enhancing a robot’s capability: Lessons from the 2015 darpa robotics [26] K. Nagasaka, H. Inoue, and M. Inaba, “Dynamic walking pattern
challenge finals,” J. Field Robot., vol. 34, no. 2, pp. 359–380, 2017. generation for a humanoid robot based on optimal gradient method,”
[2] K. Miura, M. Morisawa, F. Kanehiro, S. Kajita, K. Kaneko, and in Int. Conf. Syst. Man. Cybernet., vol. 6. IEEE, 1999, pp. 908–913.
K. Yokoi, “Human-like walking with toe supporting for humanoids,”
in IEEE/RSJ Int. Conf. Intell. Robot. Syst., 2011, pp. 4428–4435.

52

You might also like