Speed Regulation in 3D Robotic Walking through Motion Transitions

between Human-Inspired Partial Hybrid Zero Dynamics

Matthew J. Powell, Ayonga Hereid, and Aaron D. Ames
AbstractThis paper employs the Human-Inspired Control
framework in the formal design, optimization and implemen-
tation of controllers for 3D bipedal robotic walking. In this
framework, controllers drive the robot to a low-dimensional
representation, termed the partial hybrid zero dynamics, which
is shaped by the parameters of the outputs describing human
locomotion data. The main result of this paper is the use
of partial hybrid zero dynamics in an optimization problem
to compute physical constraints on the robot, without inte-
grating the dynamics of the system, and while simultaneously
yielding provably stable walking controllers for a 3D robot
model. Controllers corresponding to various walking speeds
are obtained through a second speed regulation optimization,
and formal methods are presented which provide smooth
transitions between walking speeds. These formal results are
demonstrated through simulation and utilized to obtain 3D
walking experimentally with the NAO robot.
I. INTRODUCTION
Three-dimensional bipedal robotic walking has been real-
ized experimentally by numerous robotic systems through the
use of various control schemes[1]. One of the most prevalent
control approaches leverages the Zero Moment Point (ZMP)
[2], [3], which is the control scheme included as the default
walking for the NAO
1
robot platform used as a testbed
for the controllers designed in this work. From a purely
mechanical point of view, passive walkers [4] employ an
excellent understanding of mechanics and mechanism design
to experimentally achieve robotic walking down small slopes
without the use of control. These ideas have been used to
design passivity-based control laws in 2D [5], [6]. These 2D
control laws have been extended to 3D through geometric
reduction [7], [8], yet these methods have only recently been
realized experimentally [9]. Therefore, there exists a gap
between formal methods and experimental realization for
three-dimensional robotic walking. The goal of this work is
to begin the process of bridging this gap by providing formal
results that provably result in robotic walking which can be
realized in experimentation.
The main idea behind this work is to approach 3D robotic
walking through reductions based upon virtual models and
constraints to create a low-dimensional representation of a
bipedal robot that allows formal properties of the robot
to be proven in a computationally tractable fashion. Low-
dimensional representations have been studied before; see,
M. J. Powell, A. Hereid and A. D. Ames are with the Department
of Mechanical Engineering, Texas A&M University, College Station, TX
77843, e-mail: {mjpowell,ayonga,aames}@tamu.edu
This research is supported by NASA grant NNX11AN06H, NSF grants
CNS-0953823 and CNS-1136104, and NHARP award 00512-0184-2009.
1
http://www.aldebaran-robotics.com/
for example, [10]. Similar ideas have been considered in
the past through Hybrid Zero Dynamics [11] (which has
recently been applied to 3D robots [7]) and the spring-loaded
inverted pendulum [12], [13], or SLIP model. Differing
from traditional approaches, the authors previous results
[14], [15] show that certain outputs of the human locomo-
tion control system can be represented by the solution to
an under-damped, second-order mass-spring-damper system,
and employed to achieve walking for a 2D robot model of
NAO. Novel to this paper is the application of the human-
inspired control framework to a 3D robot model of NAO,
with computation of constraints necessary to physically re-
alize walking on the actual robot.
This paper proposes a formal human-inspired optimization
(HIO) which provably results in exponentially stable bipedal
robotic walking and satises many of the physical constraints
necessary to realize the walking experimentally. Specically,
the optimization minimizes an objective function which is the
least-squares t of the output functions of the robot to the
human output data. Constraints are enforced which guarantee
that the zero dynamics surface associated with the certain
output functions is invariant through impact resulting in a
partial hybrid zero dynamics [14]. These constraints, together
with a specic choice of (linear) output functions, allow for
a closed-form approximation of the solution to the dynamics
of the robot over the course of one step, i.e., the behavior of
the robot can be determined without integrating the dynamics
of the system. This allows for the computation of physical
constraints required for experimental implementation, such
as the ZMP and friction, to be added to the HIO as constraints
and computed in a feasible time-frame (as opposed to the
time required to integrate the full dynamics of the system,
which in the case of the robot model considered in this paper
is 20-dimensional).
In addition to walking at a constant speed, formal meth-
ods are presented for obtaining walking at multiple speeds
through speed regulationin the form of another optimiza-
tion which yields controllers corresponding to a partial
hybrid zero dynamics surface for each walking speed. Motion
Transitions are constructed to smoothly connect two partial
hybrid zero dynamics surfaces. Specically, parameters of
the extended canonical walking function are obtained through
closed form expressions which satisfy PHZD at the begin-
ning and end of the step. These motion transitions allow
for seamless regulation of the robots walking speed, and
as a result, provide the ability to quickly change the robots
walking speed, as presented in the nal section on simulation
and experimental results.
II. ROBOT MODEL
The NAO robot can be modeled as a hybrid control
system:
H C
R
= (D
R
, U
R
, S
R
,
R
, f
R
, g
R
). (1)
Restrictions are imposed via control which render both feet
at throughout the gait; for non at-foot models, more
complex hybrid systems must be considered [7], [8]. The
conguration space,Q
R
, of the system is given in coordinates
by:
q = (
sa
,
sa
,
sk
,
sh
,
sh
,
nsh
,
nsh
,
nsk
,
nsa
,
nsa
)
T
,
where, as illustrated in Figure 1,
sa
,
sh
,
nsh
, and

nsa
are the stance ankle, stance hip, nonstance hip

sa

sa

sk

sh

sh

nsh

nsh

nsk

nsa

nsa
1
Fig. 1: Angle conventions
for NAO.
and nonstance ankle roll an-
gles, respectively, and
sa
,

sk
,
sh
,
nsh
,
nsk
, and

nsa
are the stance ankle,
stance knee, stance hip, non-
stance hip, nonstance knee
and nonstance ankle pitch
angles, respectively. Note
that the conguration is the
3D version of the com-
monly employed seven-link
biped model [16], [1]. With
the mass, length and inertia
properties of each link of the
robot, the Lagrangian can
be computed which, through
the Euler-Lagrange equation
(see [17]), yields the equations of motion which can be
converted to a set of rst order ordinary differential equations
(ODEs) resulting in the afne control system (f
R
, g
R
):
f
R
(q, q) =
_
q
D
1
(q)H(q, q)
_
, g
R
(q) =
_
0
D
1
(q)B(q)
_
,
with U
R
R
10
and B : Q
R
R
1010
. For the choice
of coordinates in this paper, B = I
10
. The domain species
the allowable conguration of the system, determined by a
unilateral constraint h
R
: Q
R
R; for the biped considered
in this paper, this function is the height of the non-stance
foot. In particular, the domain and guard are given by:
D
R
=
_
(q, q) TQ
R
: h
R
(q) 0
_
. (2)
S
R
=
_
(q, q) TQ
R
: h
R
(q) = 0 and dh
R
(q) q < 0
_
,
where dh
R
(q) is the Jacobian of h
R
at q. The reset map

R
: S
R
D
R
is given by:

R
(q, q) =
_

q
q

q
(q) q
_
, (3)
where
q
is the relabeling matrix which switches the stance
and non-stance legs at impact (by appropriately changing the
angles). Here,
q
determines the change in velocity due to
impact (see [18], [7] and [14]).
III. HUMAN-INSPIRED CONTROLLER DESIGN
In the authors previous work [14], it was shown that
certain outputs of human locomotion, computed from ex-
perimental locomotion data, can each be represented by a
function termed the canonical walking function (CWF):
y
H
(t, ) = e
4t
(
1
cos(
2
t) +
3
sin(
2
t)) +
5
. (4)
Motivated by the desire to obtain human-like, bipedal robotic
locomotion, the goal is to construct a controller which drives
outputs of the robot to outputs of the human. This goal is
effected formally through a control law u : TQ
R
U
R
which guarantees that y
a
(q(t)) y
d
(t) exponentially as
t , where y
a
: Q
R
R
10
is a vector of kinematics
maps on the robot representing the human outputs and y
d
:
R R
10
is a vector of canonical human functions.
With the goal of controlling the robots walking speed,
dene the relative degree 1 actual output as the velocity of
the hip and dene the the desired velocity of the hip:
y
a
1
(q, q) = p
R
hip
(q, q) = dp
R
hip
(q) q, y
d
1
= v
hip
, (5)
where p
hip
(q) is the linearized position of the hip, given by
p
R
hip
(q) = L
c
(
sa
) + L
t
(
sa

sk
). (6)
Furthermore, dene the linear (relative degree 2) actual hu-
man outputs and desired outputs represented by the walking
functions:
y
a
2,L
(q) =
_

_
m
R
nsl
(q)

sk

nsk

R
tor
(q)

sa

sh

nsh
_

_
, y
d
2,L
(t, ) =
_

_
y
H
(t,
nsl
)
y
H
(t,
sk
)
y
H
(t,
nsk
)
y
H
(t,
tor
)
y
H
(t,
sa
)
y
H
(t,
sh
)
y
H
(t,
nsh
)
_

_
(7)
where y
a
2,L
: Q
R
R
7
are the actual linear outputs of the
robot, and y
d
2,L
: R R
35
R
7
are the desired functions
for these linear outputs,
R
tor
(q) =
sa
+
sk
+
sh
, and
m
R
nsl
(q) =
sa

sk

sh

nsh
+
L
c
L
c
+ L
t

nsk
. (8)
The rst four outputs of (7) were used to obtain walking in
a 2D model of this system[15]; the three additional outputs

sa
,
sh
,
nsh
corresponding to the roll angles are novel to
this work. Due to the linear form of the outputs considered,
they can be written as:
y
a
2,L
(q) = H q (9)
for H R
710
with full row rank (where, for example, the
top row of H is obtained by taking the Jacobian of (8)).
To enforce a at non-stance foot and complete the set of
controller outputs, two nonlinear, relative degree outputs are
needed:
y
a
2,N
(q) =
_

R
x

R
y
_
, y
d
2,N
(t, ) =
_

d
x
(t,
x
)

d
y
(t,
y
)
_
, (10)
where
R
x
and
R
y
represent the roll and pitch angles of
the non-stance foot frame with respect to the ground frame.
Grouping the linear and nonlinear relative degree two outputs
results in:
y
a
2
(q) =
_
y
a
2,L
(q)
y
a
2,N
(q)
_
, y
d
2
(t, ) =
_
y
d
2,L
(t, )
y
d
2,N
(t, )
_
. (11)
where the parameters of all of the outputs are combined to
yield a single vector R
46
given by:
= (v
hip
,
nsl
,
nsk
,
sk
,
tor
,
sa
,
sh
,
nsh
,
x
,
y
).
The goal is for the outputs of the robot to agree with the
outputs of the human, motivating the nal form of the outputs
to be used in feedback linearization:
y
1
(q, q, ) = y
a
1
(q, q) v
hip
, (12)
y
2
(q, ) = y
a
2
(q) y
d
2
((q), ), (13)
where (q) =
p
R
hip
(q)p
R
hip
(q
+
)
v
hip
is a state-based param-
eterization of time with p
R
hip
(q
+
) the linearized position
of the hip of the robot at the beginning of a step. This
parameterization is important as it allows for control over
walking speed through the parameter v
hip
. These outputs can
be used to dene a human-inspired controller:
u
,
(q, q) = A(q, q)
1
__
0
L
2
f
R
y
2
(q, q)
_
(14)
+
_
L
f
Ry
1
(q, q)
2L
f
Ry
2
(q, q)
_
+
_
y
1
(q, q)

2
y
2
(q)
__
,
with control gain and decoupling matrix A given by
A(q, q) =
_
L
g
Ry
1
(q, q, )
L
g
RL
f
Ry
2
(q, q, )
_
(15)
and it follows that for a control gain > 0, the control
law u
,
: TQ
R
R
46
R
+
U
R
renders the output
exponentially stable [19].
For the hybrid control system H C
R
, the human-inspired
control law is applied to obtain the hybrid system
H
R
,
= (D
R
, S
R
,
R
, f
R
,
) (16)
with f
R
,
(q, q) = f
R
(q, q) + g
R
(q, q)u
,
(q, q). The end
result of the modeling process is a hybrid system H
R
,
that
depends on the parameters of the human inspired control
and .
Hybrid Zero Dynamics For the continuous dynamics of
the hybrid system H
R
(,)
, the controller renders the full zero
dynamics surface
FZ

=
_
(q, q) TQ
R
: y
1
(q, ) = 0, y
2
(q, q, ) = 0
9
,
L
f
Ry
2
(q, q, ) = 0
9
_
, (17)
exponentially stable (where 0
n
is a vector of n zeros). In
this work, hybrid invariance is enforced only for the relative
degree 2 outputs. The corresponding surface is referred to as
the partial zero dynamics surface (PHZD):
PZ

=
_
(q, q) TQ
R
: y
2
(q, ) = 0
9
, y
2
(q, q, ) = 0
9
_
.
(18)
Since the only output that is not included in the partial
zero dynamics surface is the output that forces the forward
hip velocity to be constant, enforcing partial hybrid zero
dynamics means, in some respect, that the velocity of the
hip is allowed to compensate for the shocks in the system
due to impact.
IV. HUMAN-INSPIRED CONTROLLER OPTIMIZATION
This section presents the main result of this paper: an op-
timization problem which yields parameters for the human-
inspired controller, u
,
, that minimize a human data-based
cost function [20] while simultaneously yielding robotic
walking in simulation and satisfying physical modeling con-
straints of the actual robot. A novel method is presented
for computing these constraints in closed form (rather than
explicitly integrating the dynamics) through the interplay
between full and partial hybrid zero dynamics.
Optimization Cost. The cost of the optimization is the least
squares t of the sagittal plane outputs to the corresponding
mean human data. The mean human data consist of discrete
times, t
H
[k], and discrete values for the output functions:
p
H
hip
[k], m
H
nsl
[k],
H
sk
[k],
H
nsk
[k], and
H
tor
[k] where here
k {1, . . . , K} N with K the number of data points.
Represent the mean human output data by y
H
i
[k] and the
canonical walking functions by y
d
i
(t,
i
) for i Output =
{hip, msl, sk, nsk, tor}; for example, y
H
msl
[k] = m
H
nsl
[k]
and y
d
msl
(t,
msl
) = m
d
nsl
(t,
nsl
). With these elements
dened, the human data based cost can be written
Cost
HD
() =
K

k=1

iOutput
_
y
H
i
[k] y
d
i
(t
H
[k],
i
)
_
2
,
(19)
which is simply the sum of squared residuals.
Partial Hybrid Zero Dynamics Constraints Following
from [14], [15], to compute the constraints needed to ensure
partial hybrid zero dynamics, the outputs and guard functions
are used to explicitly solve for the conguration of the system
() Q
R
on the guard (h
R
(()) = 0) in terms of the
parameters . In particular, let
() = q s.t. y
2
(
q
q) = 0
9
and h
R
(q) = 0 (20)
where
q
is the relabeling matrix (3). Note that multiple
solutions to () exist because y
d
2,N
(
q
q) and h
R
(q) are
nonlinear functions of all joint angles; however, restrictions
are placed on () such that only one solution corresponds
to a valid conguration. Using () allows for the explicit
solution of a point ((),

()) FZ

S
R
. In particular,
let
Y (q) =
_
dp
R
hip
(q)
dy
2
(q)
_
. (21)
and dene

() = Y
1
(())
_
v
hip
0
9
_
, (22)
where Y is invertible because of the choice of outputs. Uti-
lizing these constructions, the constraints needed for partial
hybrid zero dynamics can be written:
y
2
(()) = 0
9
, (C1)
dy
2
(
q
())
q
(())

() = 0
9
, (C2)
dh
R
(())

() < 0. (C3)
Computing approximate solutions: q
e
(t, ) and q
e
(t, )
This section utilizes the fact that the human outputs were
specically chosen to be linear in order to explicitly construct
the partial hybrid zero dynamics. Because of the specic
choice of y
a
2,L
, the following representation of the partial
zero dynamic coordinates is employed:

1
= p
R
hip
(q) =: c q, (23)

2
= y
a
1
(q, q) = p
R
hip
(q, q) =: c q.
where c R
110
is obtained from (6). This motivates the
following time-based approximation of
1
and
2
(utilizing
the solution to the inverse kinematics):

e
1
(t) := v
hip
t + p
R
hip
(
q
()), (24)

e
2
(t) := v
hip
.
These time-based approximations can be used in the partial
zero dynamics surface to obtain an approximation of the
solution for the full-order system by picking the coordinates

1
= y
a
2,L
(q) = H q, (25)

2
= L
f
Ry
a
2,L
(q, q) = H q,
with H as in (9), and dening
(
1
, ) =
_
c
H
_
1
_

1
y
d
2,L
(
1
, )
_
, (26)
(
1
, ) =
_
c
H
_
1
_
1
y
d
2,L
(1,)
1
_
. (27)
yields estimates of the rst eight angles and corresponding
velocities of the system:
q
e
1:8
(t, ) = (
e
1
(t), ), (28)
q
e
1:8
(t, ) = (
e
1
(t), )
e
2
(t). (29)
The nal four states of the system, (q
e
9:10
, q
e
9:10
) =
(
e
nsa
,
e
nsa
,

e
nsa
,
e
nsa
)
T
, are obtained through inverse kine-
matics with the assumption that the non-stance foot is parallel
to the ground throughout the step.
Model Constraints Standard methods [7] are used to
compute the ground contact forces and torques acting on
the stance foot: F
st
= (F
fx
st
, F
fy
st
, F
fz
st
, F
mx
st
, F
my
st
, F
mz
st
),
where the rst three components are the forces and the last
three components are the torques acting on the stance foot.
To prevent rotation about an edge, the following constraints
on the ground reaction moment must hold [21]:

w
f
2
F
fz
st
< F
mx
st
<
w
f
2
F
fz
st
(30)
l
h
F
fz
st
< F
my
st
< l
t
F
fz
st
, (31)
where w
f
is the width of the foot, l
t
is the length of
the toe and l
h
is the length of the heel. This condition
is known as the Zero Moment Point condition [2], [3].
Furthermore, to prevent the stance foot from slipping, the
following constraint must hold:
_
(F
fx
st
)
2
+ (F
fy
st
)
2
< F
fz
st
, (32)
where is the coefcient of static friction for the contact
between NAOs foot and the ground. Equations (30)-(32)
can be rearranged and stated in terms of inequalities of
the form C
i
(u) < 0 for i {1, . . . , 5}. Moreover, using
the approximation to the solution, (q
e
(t, ), q
e
(t, )), an
approximation the torque is computed at each time, t, over
the course of a step:
u
e
,
(t) :=u
,
(q
e
(t, ), q
e
(t, )). (33)
Therefore, the ZMP and friction constraints on the stance
foot can be stated as the constraint:
max
i{1,...,5}
max
t[0,(())]
C
i
(u
e
,
(t)) < 0, (C4)
where (()) provides an approximation of the duration of
a step (and will converge to the actual step time as ).
The nonstance foot is kept parallel to the ground via
control, through the following constraints on :

x
= 0
15
,
y
= 0
15
. (C5)
These constraints on
x
and
y
reduce the size of the
optimization search space to R
36
.
Optimization Problem Statement. The goal of human-
inspired PHZD optimization is to nd parameters

which
solve the following constrained optimization problem:

= argmin
R
46
Cost
HD
() (HIO)
s.t. (C1) (C5)
with Cost
HD
the cost given in (19). The main result of
this paper is established by combining the constructions and
results of this section with Theorem 2 of [15]. It particular,
it establishes that solving this optimization problem results
in a exponentially stable periodic orbit for H
R

,
(see [15]
for a formal denition of solutions, and the corresponding
denitions of periodic solutions and exponentially stable
periodic orbits). Furthermore, physically realistic robotic
walking can be ensured without integrating the dynamics.
V. WALKING SPEED REGULATION
The solution,

, of the optimization problem (HIO)

corresponds to a partial zero dynamics surface, PZ

, and
species controller outputs for walking at constant speed
v

hip
. However, as robotic locomotion is not always per-
formed at a constant speed, controllers which provide the
ability to smoothly transition between slow and fast walking
are essential for functional robot operation. Here, a novel
method for transitioning between these different walking
speeds is presented; specically, through the use of the
extended canonical walking function, we will connect the
Optimized Output Value Nominal Gait Human Data Mean Fitted Output
Fig. 2: Optimization for 80 values of v
hip
and comparison to human data. The tted outputs represents the canonical walking
function t to the mean human model.
PHZD surface corresponding to walking at two different
speeds. This will allow for a smooth transition between these
two walking speedsone that respects the invariance of the
PHZD surface associated to the walking at each speed.
Speed Regulation Overview. The rst step in the procedure
is to obtain optimal walking controller parameters by solving
(HIO). By denition, these parameters,

, correspond to a
local minima in Cost
HD
and satisfy the constraints (C1)
(C5). The remaining steps in the process, therefore, can
be viewed as perturbing and xing v

hip
and then solving
an optimization problem which effectively searches in a
neighborhood of

, subject to the same constraints (C1)

(C5). Choosing a small perturbation on v

hip
and using

as the seed to the speed regulation optimization results in

rapid convergence. The process is iterated using the solution
to one optimization as the seed to the next until controllers
for the desired maximum or minimum v
hip
are obtained.
Walking Speed Specication. Starting with v
0
hip
= v

hip
,
discrete walking speeds are specied via the following con-
straint
v
l1
hip
= v
l
hip
, (C6)
with l Z and the perturbation magnitude; where
is chosen based on velocity resolution and convergence
requirements (smaller leads to faster convergence in the
following optimization).
Speed Regulation Cost. The cost function corresponding to
each speed regulation step l is as follows:
Cost
SR
(,
l
) =

iOutput
_
T
0

y
d
i
(t,
i
) y
d
i
(t,
l
i
)

2
dt.
(34)
This is the integral norm of the difference between the
current controller outputs, y
d
(t, ), and the controller outputs
computed via the solution to the previous speed regulation
step, y
d
(t,
l
). As the initial seed to the speed regulation
procedure,
0
=

, is the solution to the human-inspired

PHZD optimization. This objective function serves to both
facilitate fast convergence in the following optimization (

satises (C1)(C5)) and maintain the human-like form of

the resulting walking controller outputs (

corresponds to
a local minima in Cost
HD
).
Speed Regulation Optimization. The goal of the speed
regulation optimization is to nd parameters
l+1
that solve
the following constrained optimization problem:

l+1
= argmin
R
46
Cost
SR
(,
l
) (SRO)
s.t. (C1)(C6)
with Cost
SR
the cost given in (34). The optimization (SRO)
can be iteratively solved to obtain a set of (v
MAX
hip
v
MIN
hip
)/
walking control parameters.
Extended Canonical Walking Function. It was found
in [22] that to describe more complex walking motions,
such as going up and down stairs, the canonical walking
function must be augmented to account for the role that the
environment plays on this system. Specically, the extended
canonical walking function (ECWF) is given by the time
solution to a linear mass-spring-damper system subject to
sinusoidal excitation:
y
e
H
(t,
e
i
) =e

e
i,4
t
_

e
i,1
cos(
e
i,2
t) +
e
i,3
sin(
e
i,2
t)
_
+
e
i,5
cos(
e
i,6
t) + () sin(
e
i,6
t) +
e
i,7
,
(35)
where (
e
i
) = (2
e
i,4

e
i,5

e
i,6
)/((
e
i,4
)
2
+(
e
i,2
)
2
(
e
i,6
)
2
)
and i Outputs. Note that due to the linearity of the
parameters
e
i,1
,
e
i,3
,
e
i,5
and
e
i,7
in (35), we can write:
y
e
H
(t,
e
i
) = Y
e
H
(t,
e
i,2
,
e
i,4
,
e
i,6
)
_

e
i,1

e
i,3

e
i,5

e
i,7
_

_
(36)
where Y
e
H
(t,
e
i
) R
14
only depends on the parameters

e
i,2
,
e
i,4
,
e
i,6
. The CWF can naturally be written as a
special case of the ECWF by, given parameters
i
R
5
for the CWF (4), dening
e
(
i
) := (
i
, 0, 0). Through
this embedding, we can therefore consider the same human-
inspired controller that was considered for the CWF by
replacing the CWF with the ECWF in (7). Similarly, we can
0 0.5 1 1.5
0
0.2
0.4
0.6
0.8
1
1.2
O
u
t
p
u
t
V
a
l
u
e




1 0.5 0 0.5 1 1.5
2
0
2
4
A
n
g
u
l
a
r
V
e
l
o
c
i
t
y
(
r
a
d
/
s
)




0.2 0.1 0 0.1 0.2
2
1.5
1
0.5
0
0.5
1
1.5
2




0 2 4 6 8 10
0
0.5
1
1.5
Time(s)
O
u
t
p
u
t
V
a
l
u
e


y
H,1
y
H,2
y
H,3
y
H,4
y
H,5


O1 O2 O3 O4 O5
1 0.5 0 0.5 1 1.5 2
6
4
2
0
2
4
6
Angle (rad)
A
n
g
u
l
a
r
V
e
l
o
c
i
t
y
(
r
a
d
/
s
)

sa

sk

sh

nsa

nsk

nsh
0.2 0.1 0 0.1 0.2
2
1.5
1
0.5
0
0.5
1
1.5
2
Angle (rad)

sa

sh

nsa

nsh
Fig. 3: Simulation results for steady-state walking starting from a perturbed xed point (top) and speed regulation (bottom),
showing controller outputs (left), phase portraits for pitch angles (middle) and roll angles (right).
consider the PHZD surface for the ECWF which we denote
by: PZ

e. Finally, we note that since

1
is just the linearized
position of the hip, which is used to parameterize time, we
can write the parameterized ECWF as y
e
H
(
1
,
0
1
, v
hip
,
e
i
) :=
y
e
H
(
1
0
1
v
hip
,
e
i
), which is now viewed as a function of
1
.
Motion Transitions. The advantage to the ECWF is that,
given any two PHZD surfaces these surfaces can be con-
nected with the ECWF to ensure that partial hybrid zero
dynamics is maintained, i.e., the ECWF can glue together
any two PHZD surfaces; this is not possible with the CWF
as there are not enough parameters present. To see this,
let
l1
and
l
be the parameters of the CWF associated
with walking at two different successive speeds. Associated
with these parameters are the position of the hip at the
beginning and end of a step:
0,l
1
= p
hip
(
q
(
l
)) and

f,l
1
= p
hip
((
l
)). To construct a surface connecting the
the PHZD surface associated with these two walking speeds,
consider the ECWF at the beginning of a step associated to

l1
and the end of a step associated with
l
:
y
0
i
= y
e
H
(
0,l1
1
,
0,l1
1
, v
l1
hip
,
e
(
l1
i
)) (37)
y
0
i
=
d
d
1
y
e
H
(
1
,
0,l1
1
, v
l1
hip
,
e
(
l1
i
))

1=
0,l1
1
(38)
y
f
i
= y
e
H
(
f,l
1
,
0,l
1
, v
l
hip
,
e
(
l
i
)) (39)
y
f
i
=
d
d
1
y
e
H
(
1
,
0,l
1
, v
l
hip
,
e
(
l
i
))

1=
f,l
1
(40)
for i Outputs.
The goal is to nd a parameters,
e
i
, for the ECWF such
that
e
(
l1
i
) and
e
(
l
i
) can be replaced by
e
i
in (37)-(40).
To achieve the goal of determining the parameters
e
i
, we
utilize (36) to form the following matrix:
Y =
_

_
Y
e
H
(
0,l1
1
,
0,l1
1
, v
l
hip
,
e
i,2
,
e
i,4
,
e
i,6
)
d
d1
Y
e
H
(
1
,
0,l1
1
, v
l
hip
,
e
i,2
,
e
i,4
,
e
i,6
)

1=
0,l1
1
Y
e
H
(
f,l
1
,
0,l
1
, v
l
hip
,
e
i,2
,
e
i,4
,
e
i,6
)
d
d1
Y
e
H
(
1
,
0,l
1
, v
l
hip
,
e
i,2
,
e
i,4
,
e
i,6
)

1=
f,l
1
_

_
It is easy to verify that picking
e
i,2
=
l
i,2
,
e
i,4
=
l
i,4
and

e
i,6
> 0 results in Y being nonsingular. Therefore, the nal
four parameters of
e
i
can be determined by picking:
_

e
i,1

e
i,3

e
i,5

e
i,7
_

_
= Y
1
_

_
y
0
i
y
0
i
y
f
i
y
f
i
_

_
The end result are parameters
e
i
for i Outputs. The end
result of solving for
e
in this manner is that any solution
starting in PZ

l1 which transitions through PZ

e for one
step will begin the subsequent step on PZ

l . In other words,
we will have connected the PHZD surfaces PZ

l1 and
PZ

l through PZ

e, and will therefore the control laws

developed will be valid even as the robot transitions between
different speeds. This will be veried through simulation and
experimentally in the next section.
VI. SIMULATION AND EXPERIMENTAL RESULTS
This section presents both simulation and experimental re-
sults for walking at a constant speed and transitions between
multiple walking speeds.
Simulation Results. A simulation of the hybrid system,
H
R

,
, modeling NAO is performed in which the robot
starts on the point on the guard, ((

),

(

)) and is
controlled via the human-inspired control law, u

,
with
parameters

obtained through optimization (HIO) and

= 10 as the control gain. The resulting periodic orbit for
the pitch angles and roll angles of the system are given in
Figure 3. Selected frames from one step of the simulated
walking are shown in Figure 4. Furthermore, the robot can be
started from rest, (q(0), q(0)) = (0, 0), and it will converge
to the periodic orbit imply robustness of the walking (in
simulation). Figure 4 also shows the angles of the biped in
simulation and in the experiment described later.
To demonstrate speed regulation, a simulation was con-
ducted using a series of controller parameters
l
determined
from solving the optimization problem described above for
various different choices of v
hip
between 0.14 m/s and 0.32
m/s. Transitions increased speed by 0.03 m/s every four
steps. As shown in Figure 3, the actual outputs converge to
the desired controller outputs on each step.
Experimental Results. The human inspired control approach
is implemented experimentally on the actual NAO robot via
pseudo-feedback control which uses the NAOs built-in PID
controller is to track the q(t) trajectories from simulation.
Hybrid domain switches are determined via data from the
force sensors in the feetltering of this data to effect
debouncing induces lag in the experimental system as
compared to simulation. The relative degree 2, actual control
outputs, y
a
2
, are computed from both simulation and exper-
iment data and compared against one another in Figure 3,
which shows that the experimental outputs agree closely with
the simulated outputs (with minor discrepancies in
sk
,
tor
and
sa
, which are a result of the open-loop controller).
Snapshots of the experimental walking are given with the
simulated gait in Figure 4.
Speed regulation is also implemented experimentally using
a series of controller parameters
l
determined from solv-
ing the optimization problem described above for various
different choices of v
hip
between 0.14 m/s and 0.32 m/s.
Transitions increased speed by 0.03 m/s every four steps.
Without Motion Transitions, the max achievable speed is
0.23 m/s, however, with Motion Transitions computed via
the extended canonical function (35), a max speed of 0.32
m/s is obtained (nearly a 50% increase in top speed!).
A video of the experimental walking achieve on NAO is
available online [23].
VII. CONCLUDING REMARKS
This paper presented the rst steps toward dening an
optimization problem that provably results in stable robotic
walking in 3D through the use of human output data and
controllers inspired by these data. The fundamental contri-
bution is in the form of constraints that ensure physically
realizable walking and can be enforced through solutions
obtained through the low-dimensional representation given
by partial hybrid zero dynamics. Speed regulation enables
the rapid development of walking for a variety of speeds,
and with Motion Transitions, yields 9experimentally realized
3D walking with NAO.
REFERENCES
[1] D. J. Braun, J. E. Mitchell, and M. Goldfarb, Actuated dynamic
walking in a seven-link biped robot, IEEE Trans. on Robotics, vol. 17,
no. 1, pp. 14756, Feb. 2012.
[2] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi,
and H. Hirukawa, Biped walking pattern generator allowing auxiliary
ZMP control, in IEEE/RSJ Intl. Conf. on Intelligent Robots and
Systems, Beijing, Oct. 2006, pp. 29939.
[3] M. Vukobratovi c and B. Borovac, Zero-moment pointthirty-ve
years of its life, Intl. J. of Humanoid Robotics, vol. 1, no. 1, pp.
15773, Mar. 2005.
[4] T. McGeer, Passive dynamic walking, Intl. J. of Robotics Research,
vol. 9, no. 2, pp. 6282, Apr. 1990.
[5] S. Collins, A. Ruina, R. Tedrake, and M. Wisse, Efcient bipedal
robots based on passive-dynamic walkers, Science, vol. 307, no. 5712,
pp. 10821085, Feb. 2005.
[6] M. W. Spong and F. Bullo, Controlled symmetries and passive
walking, IEEE Trans. on Automatic Control, vol. 50, no. 7, pp. 1025
31, 2005.
[7] J. W. Grizzle, C. Chevallereau, A. D. Ames, and R. W. Sinnet, 3D
bipedal robotic walking: models, feedback control, and open prob-
lems, in IFAC Symposium on Nonlinear Control Systems, Bologna,
Sep. 2010.
[8] R. W. Sinnet and A. D. Ames, 3D bipedal walking with knees and
feet: A hybrid geometric approach, in Joint 48th IEEE Conf. on
Decision and Control and 28th Chinese Control Conf., Shanghai, Dec.
2009, pp. 320813.
[9] R. W.Sinnet and A. D.Ames, Bio-inspired feedback control of three-
dimensional humanlike bipedal robots, J. of Robotics and Mechatron-
ics, vol. 24, no. 4, pp. 595601, Aug. 2012.
[10] S. Srinivasan, I. A. Raptis, and E. R. Westervelt, Low-dimensional
sagittal plane model of normal human walking, ASME J. of Biome-
chanical Eng., vol. 130, no. 5, Oct. 2008.
[11] E. R. Westervelt, J. W. Grizzle, and D. E. Koditschek, Hybrid zero
dynamics of planar biped walkers, IEEE Trans. on Automatic Control,
vol. 48, no. 1, pp. 4256, 2003.
[12] P. Holmes, R. J. Full, D. E. Koditschek, and J. Guckenheimer, The
dynamics of legged locomotion: Models, analyses, and challenges,
SIAM Review, vol. 48, no. 2, pp. 207304, Feb. 2006.
[13] I. Mordatch, M. de Lasa, and A. Hertzmann, Robust physics-based
locomotion using low-dimensional planning, ACM Trans. on Graph-
ics, vol. 29, no. 4, pp. 71:18, Jul. 2010.
[14] A. D. Ames, First steps toward automatically generating bipedal
robotic walking from human data, in Robotic Motion and Control
2011, ser. LNICS, vol. 422. Springer, 2012, pp. 89116.
[15] A. D. Ames, E. A. Cousineau, and M. J. Powell, Dynamically stable
bipedal robotic walking with NAO via human-inspired hybrid zero
dynamics, in Hybrid Systems: Computation and Control, Beijing, Apr.
2012, pp. 13544.
[16] Y. Huang, Q. Wang, B. Chen, G. Xie, and L. Wang, Modeling and
gait selection of passivity-based seven-link bipeds with dynamic series
of walking phases, Robotica, vol. 30, no. 1, pp. 3951, Jan. 2012.
[17] R. M. Murray, Z. Li, and S. S. Sastry, A Mathematical Introduction
to Robotic Manipulation. Boca Raton: CRC Press, 1994.
[18] Y. H urm uzl u and D. B. Marghitu, Rigid body collions of planar
kinematic chains with multiple contact points, Intl. J. of Robotics
Research, vol. 13, no. 1, pp. 8292, Feb. 1994.
[19] S. S. Sastry, Nonlinear Systems: Analysis, Stability and Control. New
York: Springer, 1999.
[20] A. D. Ames, R. Vasudevan, and R. Bajcsy, Human-data based cost
of bipedal robotic walking, in 14th Intl. Conf. on Hybrid Systems:
Computation and Control, Chicago, Apr. 2011, pp. 15362.
[21] C. Chevallereau, G. Bessonnet, G. Abba, and Y. Aoustin, Bipedal
Robots: Modeling, Design and Walking Synthesis. New York: Wiley-
ISTE, 2009.
[22] M. J. Powell, H. Zhao, and A. D. Ames, Motion primitives for human-
inspired bipedal robotic locomotion: Walking and stair climbing, in
IEEE Intl. Conf. Robotics and Automation, St.Paul, May 2012, pp.
543549.
[23] Speed regulation in 3D robotic walking through motion transitions,
http://youtu.be/kLakY9rWh6Y.
0 0.5 1 1.5 2 2.5
0.04
0.02
0
0.02
0.04

S
sa

E
d
sa

Ea
sa
0 0.5 1 1.5 2 2.5
0.7
0.65
0.6
0.55
0.5
0.45
0.4

S
sa

E
d
sa

Ea
sa
0 5 10
0.05
0
0.05

S
sa

E
d
sa

E
a
sa
0 5 10
1
0.8
0.6
0.4

S
sa

E
d
sa

E
a
sa
0 0.5 1 1.5 2 2.5
0.8
0.85
0.9
0.95
1
1.05
1.1

S
sk

E
d
sk

Ea
sk
0 0.5 1 1.5 2 2.5
0.5
0.4
0.3
0.2
0.1

S
sh

E
d
sh

Ea
sh
0 5 10
1
1.2
1.4
1.6

S
sk

E
d
sk

E
a
sk
0 5 10
1
0.8
0.6
0.4
0.2

S
sh

E
d
sh

E
a
sh
0 0.5 1 1.5 2 2.5
0.08
0.04
0
0.04
0.08

S
sh

E
d
sh

Ea
sh
0 0.5 1 1.5 2 2.5
0.2
0.1
0
0.1
0.2

S
nsh

E
d
nsh

Ea
nsh
0 5 10
0.1
0.05
0
0.05
0.1

S
sh

E
d
sh

E
a
sh
0 5 10
0.1
0.05
0
0.05
0.1

S
nsh

E
d
nsh

E
a
nsh
0 0.5 1 1.5 2 2.5
0.5
0.4
0.3
0.2
0.1

S
nsh

E
d
nsh

Ea
nsh
0 0.5 1 1.5 2 2.5
0.8
0.9
1
1.1
1.2
1.3

S
nsk

E
d
nsk

Ea
nsk
0 5 10
1
0.8
0.6
0.4
0.2

S
nsh

E
d
nsh

E
a
nsh
0 5 10
0.8
1
1.2
1.4
1.6
1.8

S
nsk

E
d
nsk

E
a
nsk
0 0.5 1 1.5 2 2.5
0.8
0.7
0.6
0.5
0.4

S
nsa

E
d
nsa

Ea
nsa
0 0.5 1 1.5 2 2.5
0.2
0.1
0
0.1
0.2

S
nsa

E
d
nsa

Ea
nsa
0 5 10
1
0.8
0.6
0.4

S
nsa

E
d
nsa

E
a
nsa
0 5 10
0.2
0.1
0
0.1
0.2

S
nsa

E
d
nsa

E
a
nsa
Fig. 4: Comparison of the snapshots of the actual (top row) and simulated (second row) walking gaits over one step, and
experiment and simulation for steady-state walking (left two columns) and speed regulation (right two columns): X
S
are
from simulation, X
E
d
are desired values from experiment, and X
Ea
are actual values from experiment.