Real-Time gait planning for Rh-1 humanoid robot,

using Local Axis Gait algorithm

Mario Arbulu, Carlos Balaguer
Robotics Lab, Department of System and Automation Engineering, University Carlos III of Madrid
Av. de la Universidad 30, 28911, Legans, Madrid, Spain
[email protected]
[email protected]

Abstract— It is proposed the Local Axis Gait algorithm in II. F OOT M OTION P LANNING
order to generate real-time gait patterns for a humanoid robot.
The 3D foot motion planning for the humanoid global motion A. Overview
is developed in order to walk in any surface as plane, ramp, In order to change the global position of the humanoid
climbing stairs. Furthermore, it is possible continuous change
robot, the swing foot should change its configuration from
the step length and orientation in real time. The cart-table
model is used for planning COG and ZMP motion. The proposed initial position and orientation to goal position and orientation.
algorithm takes into account physical robot constraints as joint The goal configuration is limited by the physical constraint
angles, angular velocity and torques. Torques are computed by of the robot as, angular joint limits, angular joint velocities,
Lagrange method under Lie groups. Some results are shown and angular accelerations, singularities and joint torques. The
discussed.
kinematics model using lie groups, which ∈ SE(3) [7], [8],
I. I NTRODUCTION allows to avoid internal singularities, so the problem is to
Several works about gait pattern generation have been avoid the external ones. The dynamics model use the Lagrange
developed. It is well known that the humanoid motion planning formulation and lie groups, and it is possible to compute the
uses COG and swing foot motion as reference patterns. The joint torques accurately (eq. 1 and 2).
COG motion maintain the robot stability and the swing foot    
motion allows the humanoid change its global position, so it L θ, θ̇ = K θ, θ̇ − V (θ) , (1)
is necessary to develop suitable trajectories for each humanoid
robot step. Mass distributed model developed i.e. Takanishi et. K models the potential energy effect over the body and V
al [1], Vukobratovic et. al [2], taking into account each mass models the kinetics energy over one. So the motion equation
link in order to compute the ZMP by a complex dynamic is the following 2:
model. The well known mass concentrated model i.e Kajita 
d ∂L

∂L
et. al. [3], [4], simplify the complex and non-linear humanoid − = Γ, (2)
dt ∂ θ̇ ∂θ
robot dynamics to linear one, by concentrating the overall
robot mass in its COG (center of gravity) and constraining Where Γ, is the torque due to the inertial force and gravity
the motion to any arbitrary plane. So, it is possible to develop field and it will be detailed, in the next sections. The 3D
the COG motion by following the inverted pendulum laws, in foot motion planning for carrying out the n th step can be
this paper it is used. Nevertheless, the foot motion planning modeled, as shown in Fig. 1. The swing foot moves from initial
is not clearly explained in any works [5], [6]. For that, in this configuration (position and orientation (n − 1 )th frame) to the
paper this fact will be detailed. Foot motion planning should goal configuration (position and orientation (n + 1 )th frame).
taken into account allowing the angular joint, angular velocity The input parameters are the swing foot goal configuration
and torque constraints, in specially the knee joint; furthermore, (position and orientation), defined in the appropriate local axis
avoiding collisions with support leg (internal collisions) and as: (Lx , Ly , Lz , θx , θy , θz )n+1 . It is possible to identify the
with any obstacle on walking surface (external collisions). It is parameters and frames for the n th step as follows:
obtained continuous walking motion while changing direction, P
as human natural walking. The present work is divided as Pn: World frame
following: at first, in section II, foot motion planning is : n th local configuration frame, support foot
Pn−1 th
developed. In section III, COG and ZMP motion planning : (n − 1 ) local configuration frame, swing foot
Pn+1 th
is explained. In the section IV, Lagrange formulation using : (n + 1 ) local configuration frame, swing foot
Lie groups and Screw theory is developed in order to verify pn : n th position of support foot local frame
th
the joints torque constraints. In section V, some simulation on pn−1 : (n − 1 ) position of swing foot local frame
n+1 th
Rh- 1 simulator and experimental results on Rh-1 humanoid p : (n + 1 ) position of swing foot local frame
th
platform are shown and discussed. Finally in section VI the Ln+1
x : (n + 1 ) frontal swing foot motion (x)
th
conclusions of this work are presented. Ln+1
y : (n + 1 ) lateral swing foot motion (y)

978-1-4244-1862-6/07/$25.00 © 2007 IEEE 563 HUMANOIDS’07

 th
Ln+1
z : (n + 1 ) vertical swing foot motion (z) xij : Initial position, of j-th zone
th
θxn+1 : (n + 1 ) rotation in X axis world ẋij : Initial velocity, of j-th zone
th
n+1
θy : (n + 1 ) rotation in Y axis world ẍij : Initial acceleration, of j-th zone
θzn+1 : (n + 1 )
th
rotation in Z axis world xfj : Final position, of j-th zone
ẋfj : Final velocity, of j-th zone
ẍfj : Final acceleration, of j-th zone

Hence the polynomial coefficients (eq. 5 to 10), can be de-

veloped as following from the boundary conditions described
above:

w1j = xji (5)

w2j = ẋji (6)

ẍji
w3j = (7)
2
Fig. 1. Swing foot (SF) and reference frames for 3D planning motion on
each step. In this case, the n th step planning is shown.  j j   
x −x ẋji ẍji ẋjf −ẋji ẍji
w4j = 10 fT 3 i − T2 − 2T −4 T2 − T +
B. Motion Interpolators  j j (8)
ẍf −ẍi
Polynomials have been used for approximation because they ......... 12 T
can be evaluated, differentiated, and integrated easily and
infinitely for many steps using just basic arithmetic operations    
xjf −xji ẋji ẍji ẋjf −ẋji ẍji
of addition, subtraction and multiplication. A polynomial of w5j = −15 T4 − T3 − 2T 2 +7 T3 − T2 −
order n or degree n is a function of the form:  j j
ẍ −ẍ
......... fT 2 i
n
X (9)
φ (t) = w1 + w2 .t + ... + wn .tn−1 = wk .tk−1 (3)
k=1    
n n xjf −xji ẋj ẍji ẋjf −ẋji ẍji
Where φ(t)∈ Q , with Q being the set or linear space w6j = 6 T5 − T i4 − T3 −3 T4 − T3 +
of all polynomials (eq. 3) of order n . The swing foot (SF)   (10)
ẍjf −ẍi
j
motion is planned by two fifth order interpolators (eq. 4) .......... 12 T3
one for climbing the foot (φ1 (t), eq. 4) and one for landing
the foot (φ2 (t), eq. 4), on each axis motion and orientation The continuity at the end of the climbing zone and at the
motion. These allow us to control the foot position, velocity start of the landing one should satisfy the following conditions:
and acceleration, so is minimized the foot landing reacting
force and the knee angular velocity and smooth and natural φ1 (T /2) = φ2 (T /2) (11)
joints motion are obtained.
   φ˙1 (T /2) = φ˙2 (T /2) (12)


 φ1 t, x1i , ẋ1i , ẍ1i , x1f , ẋ1f , ẍ1f ∀ 0 ≤ t < T2

φ1 (t) = w61 .t5 + w51 .t4 + w41 .t3 + w31 .t2 ...


φ¨1 (T /2) = φ¨2 (T /2)

(13)

+w21 .t + w11





φ (t) =  
φ2 t, x2i , ẋ2i , ẍ2i , x2f , ẋ2f , ẍ2f ∀ T C. Planning foot position and orientation
2 ≤t≤T



Goal foot configuration (position, P n+1 , and orientation,

φ2 (t) = w6 .t + w5 .t + w4 .t + w32 .t2 ...
2 5 2 4 2 3



n+1
θ



 +w22 .t + w12 ) are the input parameters for carrying out the next step.

 It can be done by humanoid sensors or external command.
(4) Those input parameters can be generalized in order to compute
Where the fifth order polynomial are characterized by the the n th step in real time as:
next boundary conditions on each zone; i.e. for j-th zone (j=1
for climbing, j=2 for landing): P n = P n−1 + R(θzn )T .Ln (14)
T : Step time Where,

564
Fig. 2. Swing foot (SF) motion planning changing walking direction in real
time. The n-th step trajectory is in black bold line, and the swing foot is
represented by the red bold rectangles, while the support foot is represented
by the blue bold rectangle.
Fig. 3. ZMP and COG motion planning in the n-th step.

pnx 4), allows that the ZMP trajectory to be computed by the

   n−1 
px
P n =  pny , P n−1 =  pn−1
y
, following equations:
pnz pn−1
z
Zc
ZM Px = x − ẍ (16)
g
cos(θzn ) −sin(θzn ) Lnx
   
0
R(θz ) = sin(θzn ) cos(θzn )
n T  0 , L = -(-1)n Lny 
 n 
Zc
0 0 1 Lnz ZM Py = y − ÿ (17)
g
As we see in eq. 14, the goal foot configuration depends Where x and y give us the COG trajectory (i.e. Fig. 3) at Zc
on the support foot position, because the support foot is the plane. Hence, given the reference ZMP trajectory, the COG
local axis of gait input parameters. By applying in the splines one is an inverse problem.
developed above, between local axis n-1-th to n+1-th (from
n-1 swing foot configuration to n+1 one) smooth and natural
foot motion is obtained. Some simulation results are shown
in Fig. 2. It is possible to walk changing direction and step
length in real time by using the proposed algorithm. Any
configuration motion can be obtained, but it is limited by the
physical humanoid constraints such as angular joint motion,
angular velocities and joint torques. The Lagrange method
and screws are used to compute dynamic joint effects such
as constraint torques. Furthermore, stability margin should be
taken into account, so the ZMP must be maintained in the
convex hull. The next section deals with stability humanoid
robot motion.

III. COG AND ZMP M OTION P LANNING

At the n-th step, the ZMP reference must be in the P n
point in single support and between P n and P n+1 in double Fig. 4. A Cart-table model, on x-axis.
support. Thus, we define two change foot points, at starting
n-th step and at finishing one. These points will be defined as
n
median ZMP (ZM P ). IV. DYNAMIC M ODEL
n
It is possible to generalize the ZM P configuration as A. Lagrange formulation (Lagrangian)
following (i.e. Fig. 3):
In order to compute the joint torques and dynamics con-
n n+1 straints, a dynamic model should be proposed, so the Lagrange
ZM P = R(θzn+1 )T .(L/2) (15)
formulation under the lie groups and screw theory has been
In order to define the ZMP and COG reference trajectories, developed, because it gives us a natural description of Jacobian
the cart-table model (Kajita et. al., [3]) has been used. The manipulator. The Lagrange formulation could be expressed as
control of the cart-table by the preview controler (i.e. Fig. the following:

565
    
L θ, θ̇ = K θ, θ̇ − V (θ) (18)

Where K(θ, θ̇) is the overall links kinetics energy contri-

bution and V (θ) the potential one:
  X n   1
K θ, θ̇ = Ki θ, θ̇ = θ̇T M m (θ) θ̇ (19)
i=1
2
n
X n
X
V (θ) = Vi (θ) = mi ghi (θ) (20)
i=1 i=1

With the motion equation by Γ torques being the following:

 
d ∂L ∂L
− = Γ, (21)
dt ∂ θ̇ ∂θ
Which could be expressed as the following:
 
M.m (θ) .θ̈ + N θ, θ̇ = Γ (22)

So, M .m(θ) ∈ ℜnxn is the inertial manipulator matrix and

it is defined by the operational Jacobian manipulator Jsli and
the inertial generalized matrix Mi as:
n
X
T
M.m (θ) = Jsli (θ) .Mi .Jsli (θ) (23)
i=1

The humanoid robot platform Rh-1, developed in the

Roboticslab of the University Carlos III of Madrid, which
had been used in this work, to test this algorithm; has 21
degrees of freedom, for this M .m(θ) ∈ ℜ21x21 . The angular Fig. 5. Joints torque of the support foot in the n-th step.
acceleration vector is defined by θ̈, in which has each joint
angular acceleration. Thus the joints torque without  potential
forces have been made by M .m(θ).θ̈. The N θ, θ̇ term gives The joint 9 torque limit is 44.8 Nm and the maximum torque
obtained is 12 N.m, thus its acceptable. Finally, the 12 joint
us the joints torque due to the potential forces, and is defined
limit is 17.6 Nm, the chart shown 13 N.m as maximum, it
as (eq. 24):
is near the nominal one, because the frontal ankle joint has
  ∂V less reduction respect to other ones. Furthermore, this joint
N θ, θ̇ = (24) support overall humanoid robot mass during single support
∂θ
Pn
From (20) , V (θ) = i=1 mi ghi (θ), where mi is the mass stage and it drives the frontal balance in order to maintain the
of the i-th link, g is the gravity acceleration and hi (θ) the COG ZMP inside the convex hull. We could conclude that the gait
height of the i-th link. pattern generated have satisfied the nominal torque actuators
limits, so it is validated, the gait pattern generation algorithm
B. Torque analysis in this fact. The next section, other aspects as angular range
The joints torque for the n-th step lead us to compare with and angular velocity constraints could be validated.
each joint actuator physical limit which is a dynamic constraint V. S IMULATION AND E XPERIMENTAL R ESULTS
of gait pattern. Some results are shown in Fig. 5. The support
foot joints from 8 to 12, the joint 7 is for horizontal motion A. Rh-1 humanoid robot platform
and it is not shown. Two joints drive the frontal support leg Several simulation tests have been done in Rh-1 simulator
motion: 8 (hip) and 12 (ankle). The next three joints drive software, developed under VRML environment, before to test
the sagital support leg motion 9 (hip), 10 (knee), 11 (ankle). in Rh-1 humanoid robot platform (Fig. 6), which specifications
The joints actuator torque limits are defined by the reduction are described in Table I.
gear and motor nominal torque. So, the limit nominal torque
for actuators 8, 10 and 11 is 70.4 N.m, as it is shown in the B. Gait patterns and kinematics constraints
chart that each joint never overlap the nominal torque limit; Some gait patterns should be seen in Fig. 7, taking into
it is notice that the knee joint have the highest torque, the account the angular joint range limits as see in Table I.
gait pattern generated should have in account to reduce it. The proposed gait generation algorithm allows to change

566
 TABLE I
R H -1 H UMANOID ROBOT S PECIFICATIONS

Link Number of DOF Total

Head - -
Waist 1 (Yaw) 1
Arm 4 4x2
Shoulder 1 (Pitch)
1 (Roll)
Elbow 1 (Pitch)
Wrist 1 (Roll)
Leg 6 6x2
Hip 1 (Yaw)
1 (Roll)
1 (Pitch)
Knee 1 (Pitch)
Ankle 1 (Pitch)
1 (Roll)
Total 21
Joint Human angular range Rh-1 angular range
(deg) (deg)
Head -50 to 50 (Roll) -
Fig. 6. Rh-1 humanoid robot (a) and its mechanical configuration (b).
-50 to 60 (Pitch -
-70 to 70 (Yaw) -
Arm
Shoulder -90 to 0 (Roll) -95 to 0 (Roll)
-180 to 50 (Pitch) -180 to 60 (Pitch) D. Actual results on Rh-1 humanoid robot platform
-90 to 90 (Yaw) -
Elbow -145 to 90 (Pitch) -135 to 0 (Pitch) In order to validate the proposed algorithm in the Rh-1
-90 to 90 (Yaw) -
Wrist -55 to 25 (Roll) -90 to 90 (Roll) humanoid robot, many test have been done. As observed in
-70 to 90 (Pitch) - Fig. 10, two examples of straight forward walking motion is
Hand 0 to 90 (Pitch) -16 to 60 (Pitch) shown, one in the lab (Fig. 10.a), and the other one in the hall
Waist -50 to 50 (Roll) - (Fig. 10.b); experimental walking stable motion is obtained,
-30 to 45 (Pitch) - so the algorithm is succesfully tested, up to 0.7 Km/h walking
-40 to 40 (Yaw) -45 to 45 (Yaw)
Leg
speed.
Hip -45 to 20 (Roll) -20 to 20 (Roll)
-125 to 15 (Pitch) -80 to 80 (Pitch)
-45 to 45 (Yaw) -15 to 15 (Yaw)
Knee 0 to 130 (Pitch) 0 to 100 (Pitch)
Ankle -20 to 30 (Roll) -20 to 20 (Roll)
-20 to 45 (Pitch) -25 to 25 (Pitch)

direction while the humanoid is walking, continuous step

length changing and walking in any surface. Angular velocity
constraints have been satisfied as actuator limits, thus the
smooth gait patterns could be validated; i.e the actuator knee
angular velocity limit is 8000 RPM, the gait pattern generated
is suitable for this requirement as we can see in Fig. 9.

C. VRML Rh-1 simulations

Snapshots for doing a circle and curve path on flat surface

have shown in Fig. 8. Natural and smooth change direction
could be done by the proposed algorithm, thus natural walking
is obtained as human one. Furthermore, faster walking motion Fig. 7. Gait patterns generated for continuous humanoid walking. It is
possible to generate diagonal walking (a), and any direction continuous
should be developed until 1 Km/h at any direction on flat walking (b).
surface; in another kind of surface slower walking motion by
physical constraints and stability one would be obtained.

567
 Fig. 10. Actual humanoid robot walking under LAG algorithm, (a) in the
Fig. 8. Snapshots of VRML Rh-1 simulation, for smooth naturally and laboratory, (b) in the hall.
continuous walking as figure 7.b

Currently works deal about global motion humanoid robot

control, in order to improve its walking motion. The next step,
is to make tests at uneven surfaces with an improved algorithm.
ACKNOWLEDGMENT
We would like to thank the members of the Robotics Lab of
University Carlos III of Madrid for their cooperation an sug-
gestions. We also express acknowledgment to the Humanoid
team for their support. This research is supported by the
CICYT (Comision Interministerial de Ciencia y Tecnologia),
ref. DPI2004-00325.
R EFERENCES
[1] H. Lim, Y. Kanesima, and A. Takanishi, “Online walking pattern gener-
ation for biped humanoid robot with trunk,” in Proceedings of the 2002
IEEE/RSJ Intl.Conference on Intelligent Robot and Systems, 2002.
[2] M. Vukobratovic and A. Frank, “One the gait stability of biped machines,”
Fig. 9. The angular velocity requirement by knee joint, its actuator run up IEEE Transactions on Automatic Control, 1970.
8000 RPM. [3] 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 IEEE International Conference on Robotics &
Automation, Taipei and Taiwan, September 14-19 2003, pp. 162–1626.
VI. C ONCLUSIONS [4] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Yokoi, and
H. Hirukawa, “Biped walking pattern generation by a simple 3d inverted
Foot and COG motion planning had been developed for pendulum model,” Autonomous Robots, vol. 17, 2003.
applying in real time, by satisfying the stability criterion [5] M. Morisawa, K. Harada, S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara,
on bipedal walking robots, joint angle, angular velocity and S. Nakaoka, and H. Hirukawa, “A biped pattern generation allowing
immediate modifications of foot pacement in real-time,” in Proceedings
torque conditions. Lagrange formulation, Lie groups and screw on Humanoids 2006, 2006.
theory had been employed in order to verify the joint torques [6] Q. Huang, K. Yokoi, S. Kajita, K. Kaneko, H. Arai, N. Koyachi,
in the humanoid robot, while it is walking, thus dynamics and K. Tanie, “Planning walking patterns for a biped robot,” IEEE
Transactions on Robotics and Automation, vol. 17, pp. 280–289, 2001.
constraints are obtained. Foot motion algorithm for any land- [7] M. Arbulu, J. M. Pardos, L. Cabas, P. Staroverov, D. Kaynov, C. Perez,
ing point had been developed and shown some simulations M. Rodriguez, and C. Balaguer, “Rh-0 humanoid full size robot’s control
results. Simulation results on Rh-1 simulator environment had strategy based on the lie logic technique,” in 5th IEEE-RAS International
Conference on Humanoid Robots, Tsukuba, Japan, Dec. 4-6 2005, pp.
been shown, and it allows to test the gait patterns before to 271–276.
apply it in actual humanoid robot. The proposed algorithm has [8] R. M. Murray, Z. Li, and S. S. Sastry, Mathematical Introduction To
been tested in the actual Rh-1 humanoid robot, successfully. Robotic Manipulation, C. Press, Ed. CRC Press, 1994.

568