Você está na página 1de 8
 
A Three Dimensional Foot Placement Plannerfor Locomotion in Very Rough Terrains
Ye Zhao and Luis Sentis (yezhao@utexas.edu, lsentis@austin.utexas.edu)
 Abstract
—Maneuvering through 3D structures nimbly ispivotal to the advancement of legged locomotion. However, fewmethods have been developed that can generate 3D gaits inthose terrains and fewer if none can be generalized to controldynamic maneuvers. In this study, foot placement planning fordynamic locomotion traversing irregular terrains is explored inthree dimensional space. Given boundary values of the centerof mass’ apexes during the gait, sagittal and lateral phase-planetrajectories are predicted based on multi-contact and invertedpendulum dynamics. To deal with the nonlinear dynamics of thecontact motions and their dimensionality, we plan a geometricsurface of motion beforehand and rely on numerical integrationto solve the models. In particular, we combine multi-contact andprismatic inverted pendulum models to resolve feet transitionsbetween steps, allowing to produce trajectory patterns similar tothose observed in human locomotion. Our contributions lay inthe following points: (1) the introduction of non planar surfacesto characterize the center of mass’ geometric behavior; (2) anautomatic gait planner that simultaneously resolves sagittal andlateral feet placements; (3) the introduction of multi-contactdynamics to smoothly transition between steps in the roughterrains.
I. I
NTRODUCTION
How is it that many legged animals are capable to nimblymaneuver on 3D surfaces but humanoid robots can onlyslowly walk on them? To tackle this deficiency, we aim atdeveloping new models characterizing 3D legged dynamicsand designing methods to find 3D feet placements thatachieve the desired gait regimes. To do so, in this paper wepresent a new 3D agile motion planner capable to maneuverin irregular terrains and in a natural manner. As such, thisplanner is aimed to control semi-autonomous legged robotsin realistic outdoor environments or for the analysis of humanmotion.We accomplish the 3D rough terrain capability by doingthe following: (1) we develop prismatic inverted pendulumdynamics to describe the sagittal and lateral single contactbehaviors, (2) we develop multi-contact models to describethe dynamics and internal forces of dual contact phases, (3)we introduce non-planar center of mass surfaces of motion toreduce the dimensionality of the model dynamics, (4) givendesired feet step locations in the sagittal plane and desiredapex sagittal velocities of the steps, we use numerical integra-tion to solve sagittal feet phase placements, (5) to smootherpeak velocities, we incorporate multi-contact phases andsolve for the corresponding dynamics given surface frictionconstraints, (6) we then extract time profiles of the generatedsteps and use them to search lateral feet placements that keepthe gait within velocity bounds, (7) we extract time profiles
Fig. 1.
 3D Schematic Diagram of Walking Profile:
 The center of mass geometric surface and the feet locations on the sagittal plane areprovided by the gait designer and can take arbitrary forms as long as theyare kinematically feasible. The center of mass position is
 p
com
, the centerof pressure (CoP) positions of the right and left feet are
 p
cop
(
LF 
)
 and
 p
cop
(
RF 
)
, CoM accelerations are
 a
com
, and reaction forces are
 
r
(
LF 
)
and
 f 
r
(
RF 
)
.
.of the center of mass and feet trajectories for verificationand control, (8) we apply our algorithm to the terrains withirregular profiles to demonstrate the validity of our work.One of the main characteristics of the proposed studyis its generalizing principles, such as combining variouscontact models, relying on numerical methods, solving forfeet placements in the phase plane, and maintaining centerof mass movement within velocity bounds. We show thepotential of our techniques in the generation of gait bymaneuvering nimbly in a terrain with strong height variationsusing a biped visualization environment and comparing itto the performance of a human walking. To validate theapplicability of our algorithm, our planner is tested on threedifferent challenging terrains sets. Also, we have recentlyshown extensions of some of our methods to other gaits suchas walking on vertical surfaces [1] or producing brachiationgaits [2]. Similar ideas could be used for generating gaitson quadruped robots. For instance in [3] it is shown thatcontrolling internal tensions during rough terrain walkingallows a quadruped to prevent slippage over an inclined
2012 12th IEEE-RAS International Conference on Humanoid RobotsNov.29-Dec.1, 2012. Business Innovation Center Osaka, Japan978-1-4673-1369-8/12/$31.00 ©2012 IEEE726
 
Fig. 2.
 Prediction of 3D Single Contact Behaviors:
 A prismatic inverted pendulum (i.e. one in which the height can change) is utilized to study thesagittal (a) and lateral (c) motion. In (a), the center of mass traverses the apex point while the center of mass in (c) bounces back before reaching the lateralfoot position. The phase diagrams (b) and (d) correspond to the sagittal and lateral center of mass phase behaviors given desired feet contact locations (redboxes), a desired center of mass surface of motion, and initial position and velocity conditions. The combined 3D motion is integrated in (e). If we considertiming issues on the lateral plane as discussed in Section III, we can derive two different trajectories shown in (f) and (h). (f) shows lateral CoM behaviorsgiven a fixed lateral foot placement and varying starting conditions. (h) corresponds to CoM trajectories derived given varying lateral foot placements anda fixed starting conditions. In (g), we analyze lateral CoM trajectories with one varying step transition.
surface.Our work attempts to advance the state-of-the-art in roughterrain locomotion [4], [5], [6], [7], [8] and possibly help outin other works that rely on dynamic stability [9], [10], [11],[12], [13], [14]. In particular, current bipedal methods usinglinearized assumptions cannot achieve human like speedsin rough terrains due to the simplicity of the models, andmethods relying on stability analysis require the analysis of periodic trajectories which do not apply to rough terrains. TheCapture Point method described in [8] represents a powerfulframework to plan feet placements. However, compared toour methods described here it only addresses gait generationin flat terrains. In contrast, our methods are applicable to non-flat 3D rough terrains because we do not utilize linearizedmodels.II. M
ATHEMATICAL
 D
ERIVATIONS
 A. Decoupled 3D Prismatic Inverted Pendulums Dynamics
When considering 3D locomotion, sagittal and lateralsingle contact behaviors are coupled together making thefoot placement generation a difficult task. However, with theassumption that the center of mass moves on a piecewiselinear 3D surface, the sagittal and lateral dynamics becomedecoupled and therefore can be independently solved.Using dynamic balance of moments, the difference be-tween the moments acting on the contact foot and the netinertial and gravitational moments, is zero. Therefore, for thesingle contact scenario (see Figure 1) moment balance canbe written as
 p
cop
k
 ×
r
k
 =
 p
com
×
com
 +
 M g
+
 m
com
.
 (1)where,
 k
 is the limb in contact with the ground,
 p
cop
k
 isthe limb’s center of pressure (CoP) point,
 
r
k
 is the threedimensional vector of reaction forces,
 
com
 and
 m
com
 arethe three dimensional vectors of center of mass inertial forcesand moments respectively, and
 g
 corresponds to the gravityfield. The above equation is vectorial and determines threeorthogonal moments. Force equilibrium can be formulated as
r
k
 =
 f 
com
 +
 M g
, which allows to rewrite Equation (1) as
 p
com
 p
cop
k
×
r
k
 =
m
com
.
 (2)For our prismatic inverted pendulum model we assume singlepoint mass [15], [16] and therefore inertial moments aboutthe center of mass can be ignored, i.e.
 m
com
 = 0
. As such,the above equation can be rewritten in vectorial form as
0
 
r
[
kz
]
 
r
[
ky
]
r
[
kz
]
 0
 
r
[
kx
]
r
[
ky
]
 
r
[
kx
]
 0
 p
com[
x
]
 p
cop
k
[
x
]
 p
com[
y
]
 p
cop
k
[
y
]
 p
com[
z
]
 p
cop
k
[
z
]
= 0
.
(3)Using the equalities
 f 
r
[
kx
]
 =
 Ma
com[
x
]
,
 f 
r
[
ky
]
 =
 Ma
com[
y
]
and
 
r
[
kz
]
 =
 
(
a
com[
z
]
 +
 g
)
, we can decompose the above
727
 
equation into the following three ones
a
com[
x
]
 =
 p
com[
x
]
 p
cop
k
[
x
]

a
com[
z
]
 +
 g
 p
com[
z
]
 p
cop
k
[
z
]
,
 (4)
a
com[
y
]
 =
 p
com[
y
]
 p
cop
k
[
y
]
·
a
com[
x
]
 p
com[
x
]
 p
cop
k
[
x
]
,
 (5)
a
com[
z
]
 =
 p
com[
z
]
 p
cop
k
[
z
]
·
a
com[
y
]
 p
com[
y
]
 p
cop
k
[
y
]
g.
 (6)where
 a
com
[
.
]
 represents the center of mass acceleration. Wewill use these three equations to formulate sagittal and lateraldynamic behavior.
 B. Center of Mass Geometric Surface
Equations (4) to (6) are not only nonlinear but alsomultivariate and therefore they pose a problem to solve them.To deal with this difficulty we first reduce the dimensionalityof the equations by planning a geometric surface of centerof mass behavior beforehand. In Figure 1 we depict anexample of a handmade surface. In this paper we don’texplore the making of the surfaces and assume they are givingto us. In this case, the surface is piecewise linear and itapproximately follows the contour of the terrain. Our surfacecan be expressed as
 p
com[
z
]
 =
a
1
 p
com[
x
]
 +
 b
1
, p
com
 
 P
1
a
2
 p
com[
x
]
 +
 b
2
, p
com
 
 P
2
...
a
 p
com[
x
]
 +
 b
, p
com
 
 P
(7)where,
 P
i
 represents the path of the CoM over surfacesegment
 i
. Moreover, the acceleration profile can be extractedby differentiating twice the above piecewise equation, i.e.if 
 p
com[
z
]
 =
 a
i
 p
com[
x
]
 +
 b
i
,
 then
 a
com[
z
]
 =
 a
i
 a
com[
x
]
.
 (8)Let us first solve the sagittal inverted pendulum dynam-ics. Plugging the position and acceleration dependenciesdescribed in (8) into Equation (4) we get
a
com[
x
]
 =
 p
com[
x
]
 p
cop
k
[
x
]

a
i
 a
com[
x
]
 +
 g
a
i
 p
com[
x
]
 +
 b
i
 p
cop
k
[
z
]
.
 (9)and since
 a
com[
x
]
 appears both on the left and right handsides, we can rewrite the equation as
a
com[
x
]
 =
 p
com[
x
]
 p
cop
k
[
x
]
·
g
a
i
 p
cop
k
[
x
]
 +
 b
i
 p
cop
k
[
z
]
.
 (10)The above equation represents an inverted pendulum of variable height that tracks the desired surface. Therefore wecall it the prismatic inverted pendulum model and we use itto describe single contact behaviors. Notice that by definingthe center of mass surface in Figure 1, our sagittal modelhas now become an ordinary differential equation that canbe easily solved via numerical integration.Let us now focus on the lateral single contact dynamics.Plugging the position and acceleration dependencies of Equa-tion (8) into (5), we get
a
com[
y
]
 =
 p
com[
y
]
 p
cop
k
[
y
]
·
a
com[
z
]
 p
com[
z
]
a
i
 p
cop
k
[
x
]
b
i
 .
 (11)Equation (6) can be rewritten by reorganizing terms as
a
com[
y
]
 =
 p
com[
y
]
 p
cop
k
[
y
]
 p
com[
z
]
 p
cop
k
[
z
]
a
com[
z
]
 +
 g
.
Expressing
 a
com[
z
]
 from (11) in terms of 
 a
com[
y
]
 and pluggingit above we get
a
com[
y
]
 =
 p
com[
z
]
a
i
 p
cop
k
[
x
]
b
i
a
com[
y
]
 p
com[
z
]
 p
cop
k
[
z
]
+
 p
com[
y
]
 p
cop
k
[
y
]
 p
com[
z
]
 p
cop
k
[
z
]
g.
 (12)Isolating
 a
com[
y
]
 from above, the term
 a
com[
z
]
 disappears andthe above equation can be written as
a
com[
y
]
 =
 p
com[
y
]
 p
cop
k
[
y
]
·
g
a
i
 p
cop
k
[
x
]
 +
 b
i
 p
cop
k
[
z
]
.
 (13)This result is important as it represents lateral single contactdynamics as an ordinary differential equation too, and there-fore independent of the dynamics of the sagittal dynamics.
C. Numerical Integration
Although we have found ordinary differential expressionsthat are decoupled for the sagittal and lateral planes, Equa-tions (10) and (13) are usually nonlinear in their most generalcase [17]. In the special case shown in this paper it turnsout that the dynamics become linear. However, consideringthe nonlinear case, a closed form solution of the dynamicbehavior cannot be obtained. To address this limitation, wedevelop numerical integration techniques to solve the modeldynamics.Suppose that we have a nonlinear differential equation forthe scalar variable
 x
, and with form
¨
x
 =
 f 
(
x,
 ˙
x
)
.
 (14)We assume that
 ¨
x
 is approximately constant for small in-crements of time. We discretize the trajectory,
 (
x
k
+1
,
 ˙
x
k
+1
)
,and derive Taylor expansions for a small disturbance,
 
, andfor initial conditions
 (
x
k
,
 ˙
x
k
,
 ¨
x
k
)
 to get
˙
x
k
+1
 
 ˙
x
k
 + ¨
x
k
,
 (15)
x
k
+1
 
x
k
 + ˙
x
k
 + 0
.
5 ¨
x
k
2
.
 (16)
728

Recompense a sua curiosidade

Tudo o que você quer ler.
A qualquer hora. Em qualquer lugar. Em qualquer dispositivo.
Sem compromisso. Cancele quando quiser.
576648e32a3d8b82ca71961b7a986505