You are on page 1of 15

IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO.

3, JUNE 2004 465

Simultaneous Control of Position and Orientation


for Ball-Plate Manipulation Problem Based
on Time-State Control Form
Hisashi Date, Member, IEEE, Mitsuji Sampei, Member, IEEE, Masato Ishikawa, Member, IEEE, and
Masanobu Koga, Member, IEEE

Abstract—This paper deals with the ball-plate manipulation


problem considered as a typical but complicated model of a
driftless nonholonomic system. Due to a strong nonlinearity of the
ball-plate system, a state equation of the kinematic model cannot
be transformed into a chained form, which is known to be effective
in constructing a feedback control law for some driftless nonholo- Fig. 1. Manipulation problem of a ball.
nomic systems. To address this problem, we utilize a time-state
control form, a kind of canonical form which covers a broader
class of systems than the chained form. This form is first applied a regrasping motion. This characteristic contributes to reduc-
to two separate subproblems, position control, in which the planar
tion of the number of the actuators and simplification of the
position of the ball is controlled but not the orientation, and
orientation control, in which the orientation is controlled without mechanical structure of the manipulator. We confine ourselves
changing the positional relation between the ball and the plates. It to a rolling motion in manipulation of the object hereafter, then
turns out that there exists a linearly uncontrollable subspace in the the problem can be reduced to the so-called ball-plate problem,
transformed subsystem, which turns into controllable by a change which consists of a pair of horizontal plates and a ball held be-
of coordinates. This implies that the system has the structure
tween the plates. Montana [1] derived a general geometric for-
of a system with two generators. We propose a control strategy
using iterative changes of coordinates, ensuring convergence in mulation of rolling motion between two rigid bodies, of which
the neighborhood of the origin. Finally, we unify the subproblems this paper refers to the special case—a ball and plates. This
into simultaneous control of position and orientation, i.e., the rolling contact condition is formulated as geometric velocity
whole configuration of the system. The important idea in the constraints, a well-known nonholonomic constraint. Kinematics
simultaneous control is the coordinate transformation, which
of systems under such constraints is expressed as a driftless state
enables us to avoid a singular point. Results of simulations show
that the proposed method achieve robustness to a measurement equation, which does not have any static continuous state feed-
noise and perturbation of radius of the ball. back stabilizing the origin as Brockett showed [2]. This fact has
prevented us from constructing a conventional feedback con-
Index Terms—Ball-plate manipulation, driftless system, multi-
generator system, nonholonomic constraint, time-state control troller, and gives us to use either a smooth time-varying con-
form. troller, a discontinuous time-invariant controller, or a hybrid
controller. On the other hand, nonholonomic constraint has an
interesting property, in that it does not reduce the system’s de-
I. INTRODUCTION
grees of freedom. This implies that the parameters of the system

R ECENTLY, a problem of dextrous manipulation of an ob-


ject using robot hands with multiple fingers has become of
interest for many researchers. The so-called dextrous manipula-
can be controlled by fewer inputs, i.e., fewer actuators.
The control objective in the paper is to steer the ball and the
plates to a desired position with a desired orientation, keeping
tion is to bring one grasp configuration to another configuration rolling contact without slip by giving only horizontal motion of
by manipulating fingers. When an object rolls on the surface of the plates, i.e., translation and rotation, as shown in Fig. 1. This
the fingers without slip, we can manipulate the object without problem can be split into two subproblems. We call the one po-
sition control, in which the position of the ball and that of the
Manuscript received August 17, 2002; revised February 25, 2003. This paper plates are controlled but the orientation of the ball is ignored,
was recommended for publication by Associate Editor H. Arai and Editor A. and the other orientation control, in which the orientation of
De Luca upon evaluation of the reviewers’ comments. This work was supported the ball is controlled without changing the positional relation
in part by a Grant-in-Aid for COE Research of the Ministry of Education, Sci-
ence, Sports and Culture of Japan. This paper was presented in part at the IEEE between the ball and the upper plate. Control of the whole con-
Conference on Decision and Control, Phoenix, AZ, 1999. figuration is called simultaneous control.
H. Date is with the National Defense Academy, Kanagawa 236-8686, Japan Open-loop solutions for the orientation control, including
(e-mail: date@nda.ac.jp).
M. Sampei is with the Tokyo Institute of Technology, Tokyo 152-8552, Japan. analysis of reachability, have been reported by many re-
M. Ishikawa was with the University of Tokyo, Tokyo 113-8656, Japan. He searchers. Li and Canny [3] analyzed the existence of an
is now with Kyoto University, Kyoto 611-0011, Japan. admissible path, achieving the objective based on differential
M. Koga is with the Kyushu Institute of Technology, Fukuoka 820-8502,
Japan. geometry. They also gave a path-planning algorithm capable
Digital Object Identifier 10.1109/TRA.2004.825267 of bringing the ball to the desired position and orientation in a
1042-296X/04$20.00 © 2004 IEEE
466 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

sequence of three movements. Jurdjevic studied the problem of


finding the path that minimizes the length of the curve traced
by the sphere on the fixed plane [4]. Bicchi [5] also proposed
open-loop solutions, expanding the sphere to an arbitrary
surface. Further analysis on controllability and motion planning
have been made for a general smooth surface [6] and for a poly-
hedral object [7]. Other work on open-loop control has been
given by Mukherjee [8]. As an effective open-loop approach for
trajectory generation, a notion of differential flatness has been
exploited with successful results for several nonholonomic
systems [9]. Although no such output is found for the ball-plate
system in the state of the art, Kiss et al. [10] proved that the Fig. 2. Top view of the ball and the upper plate.
system is equivalent to a Liouvillian system, a system with
partially flat output [11]. Above all, the path-planning problem be extended to a stabilizing controller by introducing high-gain
might be vulnerable to some disturbance or perturbation of feedback; however, the strategy may be very complicated and
system parameters. In order to overcome such a situation, some vulnerable to disturbance near the origin. Furthermore, robust-
sort of feedback control scheme is needed. ness to the measurement noise or parameter perturbation can be
Transformation to a kind of canonical form, such as a chained achieved by the convergent controller, as shown in a numerical
form, has been turned out to be effective in constructing a feed- example (Section IV). Therefore, we deal with only the conver-
back controller for such a class of systems. There are a lot of gent controller in the paper. The outline of the paper is as fol-
papers on the feedback-control problem using the chained form lows. In Section II, the configuration of the ball-plate system is
[12]–[16]. Unfortunately, however, it is not applicable to the introduced, and the differential equations that govern kinematics
ball-plate system. The reason is twofold. First, the chained form of rolling motion are derived. In Section III, we summarize the
has an exactly linearized subsystem and the condition of trans- time-state control form and present control strategies based on
formation is quite severe. Indeed, the system does not fulfill the the form. Next, we first apply the form to the two separate sub-
necessary and sufficient condition stated in [17]. Moreover, the problems, the position control and the orientation control. In the
condition is not satisfied even in approximation in the neigh- position control (Section IV), a basic control strategy is appli-
borhood of the origin, due to a linearly uncontrollable part in cable and global convergence is achieved. On the other hand,
the subsystem. By an analysis of the controllability, it turns out in the orientation control (Section V), we cannot immediately
the system has the structure of a two-generator system [18]. It apply the strategy of time-state control form due to the struc-
shows the complexity of the ball-plate system in view of feed- ture of the two-generator system. We cope with this difficulty
back control. Only a few results have been obtained for the sys- by iterative changes of coordinates and achieve local conver-
tems, which is not transformable to the chained form, one of gence. Finally, we unify the control problems into simultaneous
which is planar space robot [19]. As to the ball-plate problem, control of position and orientation, achieved as the result of the
Oriolo applied iterative open-loop control with piecewise con- three-input, eight-output control system in Section VI. Results
stant parameter, using a nilpotent approximation to achieve sta- of simulations for each control problem are shown, and the case
bility and some robustness [20]. with measurement noise and perturbation is also shown in the
The authors have proposed control strategies based on a time- simultaneous control.
state control form expression [21], [22], another kind of canon-
ical form for driftless nonholonomic systems, which was ap-
II. SYSTEM MODEL AND CONTROL PROBLEM
plied to the position control [23], the orientation control [24],
and the simultaneous control [25]. Its structure is similar to the A manipulation system consists of a ball, whose radius is
chained form, namely, the system is split into two subsystems. , and a pair of sufficiently wide horizontal plates. The lower
Although the chained form has a strong advantage from the plate is fixed on the ground, and the upper plate is driven with
viewpoint of control, the applicable class is limited, due to the rotation and translation, which are assigned as control inputs.
severe condition. On the other hand, the time-state control form Fig. 2 shows the relation on the ichnographical plane between
is simpler and covers a broader class of systems, in the sense that the position of the ball and that of the plates. First, we denote the
the chained form can be regarded as the special case of this form. planar position of the center of the ball by , that of the
A benefit of the form is that it can be applied even if the sub- upper plate by on the coordinate system fixed on the
system has a linearly uncontrollable part in the neighborhood of ground, and the rotational angle of the upper plate by . Next, we
the origin, which matters in the orientation control problem. take spherical coordinates on the surface of the ball, as shown
The major contribution of this paper is to provide a feed- in Fig. 3, given by angles of latitude and longitude. In Fig. 3, ,
back-based control strategy for a class of nonholonomic sys- , , and , respectively, denote the center of the ball, a point
tems which cannot be put in the chained form. The term “feed- of contact between the ball and the lower plate, the north pole,
back-based” implies that we do not focus on stabilizing the and the south pole. is the desired point of contact where the
whole system, but on achieving (local) convergence. That is, equator and the meridian intersect. is the closer point to
one state variable is driven by a feedforward input, while the between two intersections of the arc and the equator. The
others are stabilized by a feedback controller. The controller can orientation of the ball can be specified with three variables: the
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 467

as an eight-dimensional, three-input nonlinear system without a


drift term.
Our control objective is to bring the state to an arbitrary point,
but we can set the problem to bring all state variables to the
zero without loss of generality, due to the symmetry. Although
the number of inputs is only three, eight state variables can be
controlled, thanks to the characteristics of nonholonomic con-
straints. Discussions on the controllability, in the sense of exis-
tence of an admissible path, can be found in [3].

III. CONTROL STRATEGY


The ball-plate system represented in (1) belongs to a class
Fig. 3. Orientation of a ball. of input-affine nonlinear systems without a drift term. Many
kinematic model of systems under rolling constraints or sys-
north latitude ; the west longitude tems under the conservation law of angular momentum belong
of the contact point on the surface of the ball with the lower to driftless nonlinear systems. Car trailers or space robots are
plate; and , defined as crossing angle of (projection of the well-known examples. According to the Brockett’s theorem,
line onto the ground) and (a line parallel to the axis it is quite complicated to construct a feedback controller for
which passes the contact point ). Using these eight variables such a class of systems [2]. In some cases, one can transform
( , , , , , , , ), we can describe the configuration the system into a chained form, a kind of canonical form, to
of the whole system. design a feedback controller. There is much literature about
Now let us derive a kinematic equation of the rolling motion. control strategies which guarantee exponential convergence or
We assume that the ball does not slip nor twist on the lower plate, stability based on the chained form [12]–[16]. Unfortunately,
but the upper plate can pivot on the ball. When the upper plate however, the ball-plate system cannot be transformed into the
translates at velocity , the center of the ball moves at half chained form. Thus, we apply a more general canonical form, a
velocity . The ball also moves by time-state control form [21], [22], which is briefly summarized
rotation of the upper plate. A velocity of the ball in such a case in the following.
is in proportion to the distance between the ball and the center
A. Time-State Control Form
of the upper plate
We consider a nonlinear state equation with states and
inputs as a general model of driftless nonholonomic systems

(2)
Behavior of the orientation of the ball is derived by straight-
forward but clumsy manipulation, according to Montana [1], of where is a state vector, are vector
which the equivalent model can be found in [3], [5], and [14] functions of which are linearly independent at the origin, and
is an input vector with . As Brockett
showed, the origin of these systems cannot be stabilized by any
kind of continuous static state feedback [2]. In order to con-
trol such systems by feedback, we need either a smooth time-
varying controller, a discontinuous time-invariant controller, or
a hybrid controller. The time-state control form can be classified
into the hybrid controller.
Now we transform the general expression (2) into the time-
state control form. Under a certain condition, there exists a co-
ordinate transformation

(3)

and an input transformation


(1) (4)
such that the system (2) can be split into the following two equa-
tions:

(5)
Taking ( , , , , , , , ) as a state, and ( , , ) as
control inputs, we can regard this rolling contact condition itself (6)
468 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

where is a new state, is a new time scale, the backward time scale which monotonically increases
mapping is a (local) diffeomorphism, ’s are scalar functions while monotonically decreases, and then the state control
of and , and ’s are vector functions only of (not con- part (5) is rewritten as the backward state control part
taining ) where and are linearly
independent. We call (5) a state-control part, and (6) a time-con- (11)
trol part. This form of system expression is called a time-state
control form. The necessary and sufficient conditions for the ex- Since the linear approximation of this equation has the same
istence of the transformations (3) and (4) are given as follows. controllability as (10), there also exist differentiable functions
Theorem 1 ([26], [27]): For the nonholonomic system (2), ’s such that a feedback sta-
the following are equivalent. bilizes the subsystem (11) as long as monotonically increases,
i.e., monotonically decreases. Hence, we can bring and
1) There exist coordinate and input transformations (3), (4)
sufficiently close to zero by alternating the direction of the time
such that the system (2) is transformed into a time-state
scale with corresponding feedbacks ’s and ’s. These feed-
control form (5), (6).
backs can be designed according to conventional control theory,
2) There exists a time-scale vector field such that
such as the linear quadratic (LQ) optimal method, or a more ad-
vanced one, such as the adaptive or robust method.
(7) Stability of the state control part was discussed in [28] with
two Lyapunov functions. For precise discussion in later sections,
where is a Lie bracket. we show a control strategy which ensures exponential conver-
3) There exists a time-scale vector field and scalar valued gence of the whole states. We use the fact that the state control
functions ’s such that part (5) can be (locally) exponentially stable via the feedback
. Then for an initial time ,
.. .. there exist , , and such that
. .
(8)
(12)
where is a value of at a time . The same argument holds
for the backward state control part (11) with the feedback
linearly independent. (9) . Then there exist , , and
such that
The proof is stated in [27].
Remark 1: In order to find the coordinate and input transfor-
mations, we need to find a suitable time-scale vector field first. (13)
In some cases, it can be found in the following way. If one vector The above argument is summarized as follows.
field in (2) has one constant element, the corresponding coor- Control Strategy 1:
dinate is a candidate for , and the vector field is a candi- Step 1) Increase the new time scale by giving
date for . Once an appropriate time-scale vector field is found, in the time control part (6), and stabilize the state
the transformations can be obtained systematically. For more control part (5) until a certain time with the
detail and more general cases, see [27]. feedback . From (12),
We should note that the time-state control form itself does not the following holds at the time :
refer to the controllability of the state control part. Therefore,
the state control part is not always controllable in a linear or (14)
nonlinear sense. The time-state control form is equivalent to the Step 2) Decrease the time scale , i.e., increase the backward
chained form if the state control part (5) is exactly linearizable. time scale by giving , and stabilize the
In this sense, the time-state control form covers a broader class backward state control part (11) with the feedback
of systems. Even if the state-control part is not exactly lineariz- until the time scale
able, we can construct a feedback-based controller as follows. becomes zero. If satisfies
If a first-order approximation around the origin of the state , then exponential stability for the backward state
control part (5) control part holds as
(15)

(10) Step 3) Repeat the above two steps until all the states, and
, are sufficiently close to zero.
is linearly controllable, there exist differentiable functions This control strategy guarantees exponential convergence by
’s such that a feedback putting an additional condition on (traveling length of ).
stabilizes the subsystem (5) as long as the time scale mono- Lemma 1: is exponentially convergent if is long enough
tonically increases [2]. In the time-control part (6), we can to satisfy the following for some constant :
freely increase or decrease the time scale by setting the
appropriate sign of the input . It is expedient to introduce a (16)
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 469

Then, times repeat of the above two steps yields

Remark 2: Note that the above condition implies that the


time scale , one of the state variables in the original system,
may have to travel for a long distance, even if the state error is
small, as long as we use a fixed pair of feedback and . In
this sense, the proposed controller does not stabilize the whole
system in the Lyapunov sense. However, if the initial state is suf-
ficiently close to the origin, we can choose another pair of higher
gain feedbacks and , such that (16) would hold for arbitrarily
small , i.e., we can make and arbitrarily large. Hence,
all the states can be stabilized by changing the feedbacks and
to higher gain step-by-step. In this paper, however, we will
consider only a convergent controller, namely, a fixed pair of
feedback, since the stabilizing controller will be quite compli-
Fig. 4. Time response of the rotation angle of the upper plate  as the new time
cated and weak to disturbance because of high-gain feedback. scale  .
In addition, if we assume that the time scale can be precisely
controlled by the feedforward input, any other error due to dis-
turbance or system perturbation can be counteracted by the ro-
bust nature of the feedback controller in the state control part.
In the following, the time-state control form is applied to the
ball-plate system. We first apply the control strategy to the po-
sition control and the orientation control separately.

IV. POSITION CONTROL


The objective of the position control is to bring the position of
the ball , and the position and the rotational angle of the
upper plate ( , , ) to a desired point ignoring the orientation
of the ball. We take the origin as a desired point without loss of
generality.

A. Transformation Into Time-State Control Form


First, we extract a kinematic equation concerning the position Fig. 5. Other controlled variables u , v , u , and v .
control from the system (1) as follows:
Since the state control part (19) is a controllable linear system of
, the origin can be globally exponentially convergent via static
(17) state feedback, according to Control Strategy 1.

B. Simulation
Figs. 4–6 show results of a simulation. Initial values of states
The system (17) is a five-dimensional, three-input driftless non- are and feedback con-
linear system. Next, we transform the system into the time-state troller for the state control part (19) is designed based on LQ
control form. Here, we choose as a new time scale (see Re- optimization with weight matrices and
mark 1) and apply the following coordinate and input transfor- . Fig. 4 is a response of the new time scale
mations: versus the real time . The time scale monotonically
increases to 5.0 s while all other states converge to zero (Fig. 5).
(18) After 5.0 s, the time scale turns back and the backward stabi-
lization is applied. The same step repeats until 20.0 s, when all
state variables settle sufficiently near zero. The trajectory of the
Then, the system is transformed into the time-state control form ball and the upper plate are shown in Fig. 6.

V. ORIENTATION CONTROL
(19)
When the position control is finished, the ball and the upper
plate have the desired position, but the orientation of the ball
is uncontrolled. The objective of the orientation control is to
(20) bring the orientation of the ball ( , , ) to a desired point.
470 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

Here we see an obstruction. A linear approximation of the state


control part

(25)

is not linearly controllable, though the original system (21) is


nonlinearly controllable.
This fetter is due to a shortage of Lie brackets in the context
of nonlinear controllability. In the chained form or the time-state
control form, the controllability distribution is spanned by a se-
ries of Lie brackets with respect to a specified vector field in
the form of .
In fact, the state-control part (23) is also nonlinearly uncontrol-
lable, as well as the linear uncontrollability (see the Appendix).
Interestingly, the uncontrollable part turns into
Fig. 6. Trajectory of the ball and the upper plate. controllable if we choose another coordinate transformation,
as stated in the next subsection. This implies that the uncon-
trollable part is spanned by another series of Lie brackets with
In order to control the orientation, translation of the upper plate
respect to another vector field . In general, a system whose
is sufficient as control inputs [5]. If the upper plate
controllability distribution is spanned by two different series of
moves without rotation, the ball follows the trajectory of the
upper plate, satisfying , . There- Lie brackets is called a two-generator system [18].
In the following, we propose an extended control strategy
fore, we ignore the existence of the upper plate and consider the
using iteration of coordinates changes.
translation of the ball ( , ) as control inputs. We also assume
that the initial value of the position of the ball ( , ) is zero. B. Extended Strategy for Orientation Control
Thus, the control objective is to bring ( , , , , ) to zero.
We denote the controllable part by and the
A. Transformation Into Time-State Control Form uncontrollable part by . Then (23) is described by
A kinematic equation concerning the orientation control is
extracted from the system (1) as follows: (26)

and

(27)

Note that the right-hand side of the subsystem (26) does not con-
(21) tain the uncontrollable variable . Then the origin of the sub-
It should be noted that there is no spin between the ball and both system (26) can be (locally) exponentially stable via a linear
of the plates because the upper plate does not rotate. state feedback as long as the time scale monoton-
The system (21), a five-dimensional, two-input driftless non- ically increases, because the linear approximation around the
linear system, does not satisfy necessary and sufficient condi- origin
tion for the chained form (see the Appendix for details). But we
can transform it into the time-state control form as follows. We
chose as a new time scale and apply coordinate and input (28)
transformation as follows:
is a linear controllable system. Thus, there exist , , and ,
satisfying the inequality (12). Clearly, there also exist , ,
(22)
and for the backward state control part, satisfying (13) with
a feedback .
Then, the system is transformed into the time-state control form On the other hand, the dynamics of depends only on . Then
the closed-loop system (27) with the same input can
be expressed as
(23)
(29)

where is a scalar function of a vector . Note that the first-


(24) order approximation of vanishes at the origin.
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 471

Lemma 5: Let and , respectively, be values of and


after the change of coordinates in Lemma 4. Then the following
holds:

(32)
(33)

See the Appendix for the proof.


On the basis of the above arguments, an outline of the ex-
tended control strategy is summarized as follows.
Fig. 7. Coordinates change on the ball such that v is 0. Control Strategy 2:
Step 1) Stabilize the controllable part according to Control
This implies that the dynamics is slow if is sufficiently close Strategy 1 until is sufficiently close to zero. will
to zero. The following lemma shows that stays near the origin have small steady error at the end of this step.
with small steady error during stabilization of . Step 2) Change the coordinates so that is zero. Then the
Lemma 2: in (29) is uniformly bounded as long as the time steady error left in the uncontrollable part is put in
scale monotonically increases, if the initial value of is suf- the controllable part .
ficiently small, so that for given , , Step 3) Repeat Step 1) and Step 2) until both and are
is satisfied. Then the boundary is given as sufficiently close to zero.
(30) After each coordinates change, will again have a steady
error through the stabilization of controllable part; however, if it
where , , and are the initial values of , , and , respec- is smaller than the previous step, the whole state will converge to
tively. See the Appendix for the proof. zero by repeating the steps in Control Strategy 2. The following
It is easy to see the claim of this lemma is also applicable to theorem gives a sufficient condition for stability, showing that it
the case of the backward state control part. Next, we show that is convergent, at least in the neighborhood of origin.
is still bounded through alternate stabilization via feedback Theorem 2: Assume that the initial value for is sufficiently
and , as in Control Strategy 1. small so that Lemma 3 holds. Then, all state variables exponen-
Lemma 3: Assume that the initial value of is sufficiently tially converge to zero according to Control Strategy 2 if the
small so that Lemma 1 holds. Then is uniformly bounded following is satisfied:
during stabilization of in Control Strategy 1. The boundary
is given as
(34)
(31)
is the constant smaller than one which governs the ratio of
where and other no- the convergence in (16), is the number repetition of a pair of
tations are same as in Control Strategy 1. See the Appendix for steps in Control Strategy 1, and is that in Lemma 3. See the
the proof. Appendix for the proof.
In the following, we show that the steady error remaining in Remark 3: The condition of exponential convergence (34)
can be attenuated by applying a change of coordinates. holds if: a) is sufficiently smaller than 1; b) the number the
Lemma 4: There exist a change of coordinates which annihi- of switchings is sufficiently large; and c) the initial error on
lates the uncontrollable part without moving the origin of all and are sufficiently small, so that would be small. Espe-
the coordinates. cially the condition for initial error is quite severe because eval-
Proof: At each end of stabilization of , time scale is uation of the convergence is based on the Euclid norm of the
always zero. Hence, the uncontrollable part van- state vector. However, the proposed strategy works for much a
ishes if is zero. By arbitrariness of the coordinate system, broader region of initial value, as demonstrated by numerical
we can choose such coordinates, which are geometrically de- examples afterward.
termined as follows. It corresponds to rotation of the co-
ordinates around the line so that the meridian (great circle C. Simulation
which passes and ) passes through the current point of con- Figs. 8–14 show the results of a simulation with snapshots of
tact (Fig. 7). Note that rotating coordinates affects animation. The radius of the ball is 0.05 and the initial value of
(skew angle between coordinates and coordinates on orientation is rad. Note that the
the ground). In order to preserve the origin of this variable, initial value of is set to zero because of the arbitrariness of
coordinates should rotate around their origin. Hence, we can an- the coordinate system, as mentioned in Lemma 4. Additionally,
nihilate the uncontrollable part . by the assumption, the initial value of positional parameters ,
By this coordinates change, steady error remaining in in , , and are set to zero (Fig. 8). Stabilizing controller for
the original coordinates appears in the controllable part. The the approximated controllable part (28) is designed according
following lemma gives an evaluation of state variables through to LQ optimization, with weight matrices
this operation. and .
472 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

Fig. 8. t = 0:0 s: initial configuration.

Fig. 11. Time response of new time scale v .

Fig. 9. t = 20 0
: s: end of stabilization of controllable part before coordinates
change. An error can be seen remaining.

Fig. 12. Time responses of parameters for the orientation u , v , .

Fig. 10. t = 30 s: final configuration after stabilization in the new coordinates.

Figs. 11–13 are the time responses of the state variables. The
coordinates change occurs at 20 s (Fig. 9). The values of
and after 20 s are ones in the new coordinates. Fig. 11 is
the time response of the new time scale . Time responses of
parameters for the orientation , , are shown in Fig. 12.
Controllable variables and are stabilized to zero, while ig-
nored variable moves along with the new time scale with an
error at 20 s. By the coordinates change, the values of and
are interchanged. This indicates that the error remaining in the
previously uncontrollable part is put in the controllable sub-
space. After attenuating the error in the controllable part in the Fig. 13. Time response of ignored variable v +v .
new coordinates from 20 to 30 s, all the variables finally go to
zero. Fig. 13 shows the time response of the ignored (uncontrol- VI. SIMULTANEOUS CONTROL
lable) variable . It has steady error until 20 s; however, it
jumps to almost zero through the coordinates change. Fig. 14 is A. Avoidance of a Singular Point
the locus of the ball on the lower plate. Direction of the new time Now we are in a position to combine the subproblems into si-
scale turns counterclockwise approximately 90 , since co- multaneous control. The important idea is unification of the time
ordinates rotate around its origin by the coordinates change. scales in the time-state control form for both the position control
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 473

Fig. 16. Offset between the ball and the center of the upper plate. (left) No
offset. (right) Offset of length l.

Fig. 14. Trajectory of the ball. (a) Before coordinates change. (b) After
coordinates change. Fig. 17. Top view of desired position of the virtual center and the ball.

of the upper plate converges to a line (Fig. 17).


Moreover, (time scale in the orientation control) moves in
proportion to (the time scale in position control) in a manifold
. Hence, two time scales in both the position control and
the orientation control can be unified. Therefore, we take as
the new time scale in the simultaneous control. Finally, we have
the following eight-dimensional, three-input driftless nonlinear
Fig. 15. Virtual center of rotation on the upper plate. system as a state equation of the simultaneous control:

and the orientation control. In order to control the position of the


upper plate and that of the ball, rotation of the upper plate is
needed as the control input. However, at the origin, where planar
position of the ball and the center of the upper plate coincide,
the ball does not move, even if the upper plates rotate on the as-
sumption that the upper plate can spin on the ball. This situation
can be regarded as a singular point.
Note that the upper plate can rotate around any point. Then we
can consider a virtual center of rotation with an offset
of length from the original center (Fig. 15) by the following
input transformation:

By taking the offset, the virtual center and the ball keep a dis-
tance in the neighborhood of the origin, and hence, the singular
point is avoided (Fig. 16). Furthermore, we apply the following
coordinates transformation for the state of the system (1) into
: (36)

B. Transformation Into Time-State Control Form


We shall transform the system (36) into the time-state control
(35) form, as in the previous sections. Since we have already deter-
mined to take as the new time scale , we take a new state
Since we can take the virtual center of rotation anywhere on and inputs , , and as follows:
the upper plate, the control objective can be rewritten to bring
the new states to zero without loss of generality.
By this coordinates transformation, when converges to zero,
the ball converges to the origin, while the virtual center (37)
474 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

By these transformations, the system (36) is transformed into


time-state control form

Fig. 18. t = 0:0 s: initial configuration.

(38)
Here, we encounter the obstruction that an approximation of
the state control part (38) is not a linearly controllable system
Fig. 19. t = 10:0 s: end of stabilization before coordinates change. Planar
position is close to the origin, while large error remains in v .

(39)
This disadvantage is essentially equivalent to what we illus-
trated in Section V, and we can overcome it in the same manner.
By taking

the same discussions on the convergence immediately hold. The Fig. 20. t = 20:0 s: end of stabilization after first coordinates change.
uncontrollable part will not move so much because
its pole at the origin is zero, then we neglect this variable first,
and stabilize the rest states , whose linear approximation is a
controllable linear system, as follows:

C. Simulation Fig. 21. t = 25:0 s: end of stabilization after second coordinates change.
The radius of the ball is 0.05, and the offset between
the virtual center and original center of the upper plate is indicated. Initial values are
0.05. The state feedback for the approximated controllable (Fig. 18).
part (Section VI-B) is designed based on LQ optimization Figs. 22–25 show time responses of controlled parameters.
with weight matrices and The coordinates change occurs twice at s and
. Figs. 18–21 show snapshots of anima- s. Value after each coordinates change is that in the new co-
tion with the virtual center of rotation of the upper plate ordinates. The planar position of the ball and that of the
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 475

Fig. 22. ^ , and v^ .


Time responses of positional parameters u , v , u Fig. 24. Neglected state variable v + v .

Fig. 23. Time responses of parameters for the orientation u , v , .


Fig. 25. _ , u = v_ , and u = _ .
Time response of the control inputs u = u

virtual center of rotation on the upper plate are shown in


Fig. 22. The orientation parameters are shown in Fig. 23. Until
the first coordinates change at s (Fig. 19), all param-
eters except converge sufficiently close to zero. By the first
coordinates change at s, the error remaining in is
put to , which is controllable. According to the coordinates
change stated in Lemma 4, the coordinates rotate counter-
clockwise approximately 90 . The planar trajectory of the ball
and that of the virtual center of rotation are shown in Figs. 26
and 27, respectively. Then, all the variables are brought much
closer to the origin than the first time, with smaller steady error
in the ignored variable (Fig. 24). By the second coordi-
nates change, the error in almost vanishes (Fig. 24), and
the stabilization brings all variables sufficiently close to zero at
s (Fig. 21).

D. Simulation With Disturbance and Perturbation


In order to show the robustness to the measurement noise Fig. 26. Trajectory of the ball. (a) Before coordinates change. (b) After first
coordinates change. (c) After second coordinates change.
or perturbation of the proposed controller, the following sim-
ulations are performed. The radius of the ball is 50% larger
than the nominal value, whereas the controller is the same as controller is disturbed by Gaussian noise with zero mean and
the nominal case. In addition, the state value measured by the dispersion of 0.002. Any other conditions, such as initial value
476 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

Fig. 30. Neglected state variable v + v : perturbed case.

Fig. 27. Trajectory of the virtual center of the upper plate. (a) Before
coordinates change. (b) After first coordinates change. (c) After second
coordinates change.

_ , u = v_ , and u = _ :
Fig. 31. Time response of the control inputs u = u
perturbed case.

Fig. 28. ^ , and v^ : perturbed


Time responses of positional parameters u , v , u
case.

Fig. 32. Trajectory of the ball (perturbed case). (a) Before coordinates change.
(b) After first coordinates change. (c) After second coordinates change.
Fig. 29. Time responses of parameters for the orientation u , v , : perturbed
case.
the feedback control input is disturbed by the noise,
or the end condition, are the same as the nominal case. Com- as shown in Fig. 31, the control objective is accomplished with
pare the results in Figs. 28–33 with the nominal case. Although aslight increase of switching times.
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 477

Fig. 33. Trajectory of the virtual center of the upper plate (perturbed case). (a)
Before coordinates change. (b) After first coordinates change. (c) After second
coordinates change.

Controllability in the sense of nonlinear can be verified by


VII. CONCLUSION . The
necessary and sufficient condition for the chained form is
In this paper, we addressed a study of the feedback-based con-
not satisfied [17]. The nonlinear uncontrollability of the state
trol problem for the ball-plate system, which is a well known,
control part (23) can be verified by producing the following
but complicated, example of nonholonomic systems. Owing to
Lie brackets:
the strong nonlinearity of the ball-plate system, the state equa-
tion of the system cannot be transformed into or approximated
to the chained form, which is known to be effective for some
classes of driftless nonholonomic systems.
To this problem, we applied the time-state control form, APPENDIX II
which covers a broader class of systems that permits nonlin- PROOF OF LEMMA 2
earity and uncontrollability in the transformed subsystem. We The solution of the differential (29) is given by
first split the problem into two primitive subproblems, position
control and orientation control, and applied the time-state
control form. It turned out that an intrinsic difficulty lies in
the orientation control, due to the structure of a multigenerator Since the first-order derivative of at the origin is zero, for
system. Therefore, we extended the strategy with iteration of given , there exists such that
coordinates changes that changes the structure of controlla- implies . This inequality holds for all if
bility, with a condition for convergence in the neighborhood satisfies . By exponential stability of the
of origin. Finally, we unified the subproblems into the si- closed-loop system for (26), we have the explicit expression for
multaneous control. The important idea was the coordinate evaluation of the integral as
transformation, which enables us to avoid a singular point and
to unify the different time scales in the two subproblems.
The effectiveness of the proposed feedback-based controller
was demonstrated by simulations in the presence of disturbance
and perturbation.
Thus, we obtain
APPENDIX I
LIE ALGEBRA IN THE ORIENTATION CONTROL PROBLEM

By straightforward calculation of Lie brackets in (21), we


have the following:
478 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 20, NO. 3, JUNE 2004

APPENDIX III holds. Hence, we finally


PROOF OF LEMMA 3 obtain
Let and be the value of and at the th switching of the
time scale ( implies the initial value). At the first switching

holds from (14) and (30). In general, transition of the states


through the forward stabilization is evaluated as

APPENDIX V
and so is the case of the backward stabilization from (15)
PROOF OF THEOREM 1
Let suffixes , , , and denote “before stabiliza-
tion,” “after stabilization,” “before coordinates change,” and
“after coordinates change,” respectively. Assuming that the
After a pair of forward and backward stabilizations, the fol- number of switching of time scale is , from Lemma 3, we
lowing holds: have the following transition of the states through stabilization,
according to Control Strategy 1:

(41)
Then, infinite times of switching yields
From Lemma 4, the following holds through the coordinates
change:

(42)
Combining (41) and (42) yields
APPENDIX IV (43)

PROOF OF LEMMA 4 where the sign of inequality denotes comparison between each
Let , , , , denote values of , , , , in element. Then, and exponentially converge to zero if
the new coordinates. Since the coordinates change on the ball eigenvalues of the matrix
( coordinates) is rotation around the line (Fig. 7), the
following is satisfied from geometrical relation:
satisfy . This yields the inequality (34).

REFERENCES
This equation implies
[1] D. J. Montana, “The kinematics of contact and grasp,” Int. J. Robot. Res.,
vol. 7, no. 3, pp. 17–32, 1988.
[2] R. W. Brockett, R. S. Millman, and H. J. Sussmann, Eds., Asymptotic
Stability and Feedback Stabilization. Boston, MA: Birkhäuser, 1983,
vol. 27, Differential Geometric Control Theory, pp. 181–191.
[3] Z. Li and J. Canny, “Motion of two rigid bodies with rolling constraint,”
By the above change, coordinates rotate on the ground. IEEE Trans. Robot. Automat., vol. 6, pp. 62–72, Feb. 1990.
This causes change to , the skew angle between coordi- [4] V. Jurdjevic, “The geometry of the plate-ball problem,” Archives Ra-
tional Mech. Anal., vol. 124, no. 4, pp. 305–328, 1993.
nates and the ground coordinates. In order to preserve the origin [5] A. Bicchi and R. Sorrentino, “Dexterous manipulation through rolling,”
of all the coordinates, we need to compensate the change of by in Proc. IEEE Int. Conf. Robotics and Automation, 1996, pp. 452–457.
rotating coordinates around its origin. Taking the rotational [6] A. Marigo and A. Bicchi, “Rolling bodies with regular surface: Control-
lability theory and applications,” IEEE Trans. Automat. Contr., vol. 45,
angle , we obtain pp. 1586–1599, Sept. 2000.
[7] M. Ceccarelli, A. Marigo, S. Piccinocchi, and A. Bicchi, “Planning mo-
(40) tions of polyhedral parts by rolling,” Algorithmica, vol. 26, no. 4, pp.
560–576, 2000.
[8] R. Mukherjee, M. Minor, and J. Pukrshpan, “Simple motion planning
Note that we operate this coordinates change only after strategies for spherobot: A spherical mobile robot,” in Proc. 38th IEEE
the backward stabilization of , when is zero. Then Conf. Decision and Control, 1999, pp. 2132–2137.
DATE et al.: SIMULTANEOUS CONTROL OF POSITION AND ORIENTATION FOR BALL-PLATE MANIPULATION PROBLEM 479

[9] M. Fliess, J. Levine, P. Martin, and P. Rouchon, “Flatness and defect of Hisashi Date (S’99–M’03) received the M.E. and
nonlinear systems: Introductory theory and examples,” Int. J. Control, Ph.D. degrees in control engineering from the Tokyo
vol. 61, no. 6, pp. 1327–1361, 1995. Institute of Technology, Tokyo, Japan, in 2000 and
[10] B. Kiss, J. Levine, and B. Lantos, “On motion planning for robotic ma- 2003, respectively.
nipulation with permanent rolling contacts,” Int. J. Robot. Res., vol. 21, Since 2003, he has been a Research Associate
no. 5, pp. 443–461, 2002. in the Department of Computer Science, School
[11] A. Chelouah, “Extensions of differential flat fields and Liouvillian of Electrical and Computer Engineering, National
systems,” in Proc. 36th IEEE Conf. Decision and Control, 1997, pp. Defense Academy, Kanagawa, Japan. His current
4268–4273. research interests are in the field of nonlinear control
[12] G. Lafferriere and H. Sussmann, “Motion planning for controllable sys- theory, especially in nonholonomic systems, and
tems without drift,” in Proc. IEEE Int. Conf. Robotics and Automation, their application to robotic systems.
1991, pp. 1148–1153.
[13] R. M. Murray and S. S. Sastry, “Nonholonomic motion planning:
Steering using sinusoids,” IEEE Trans. Automat. Contr., vol. 38, pp.
700–716, May 1993. Mitsuji Sampei (M’84) received the B.Eng.,
[14] R. M. Murray, Z. Li, and S. S. Sastry, A Mathematical Introduction to M.Eng., and Dr.Eng. degrees in control engineering
Robotic Manipulation. Boca Raton, FL: CRC Press, 1994. from the Tokyo Institute of Technology, Tokyo,
[15] O. J. Sordalen and O. Egeland, “Exponential stabilization of nonholo- Japan, in 1983, 1985, and 1987, respectively.
nomic chained systems,” IEEE Trans. Automat. Contr., vol. 40, pp. From 1987 to 1991, he was a Research Associate
35–49, Jan. 1995. in the Department of Electrical and Electronics
[16] R. T. McCloskey and R. M. Murray, “Exponential stabilization of Engineering, Chiba University, Chiba, Japan. From
driftless nonlinear control,” IEEE Trans. Automat. Contr., vol. 42, pp. 1991 to 1993, he was an Associate Professor in the
614–628, May 1997. same department. From 1993 to 1999, he was an As-
[17] R. M. Murray, “Nilpotent basis for a class of nonintegrable distributions sociate Professor in the Department of Mechanical
with applications to trajectory generation for nonholonomic systems,” and Environmental Informatics, Tokyo Institute of
Math. Control, Signals, Syst., vol. 1, no. 7, pp. 58–74, 1994. Technology, and since 1999, has been a Professor in the same department. His
[18] M. Ishikawa, “Symmetric affine system with two generators: Topology current research interests are in the field of nonlinear system theory, including
and discontinuous feedback control,” in Proc. 5th IFAC Symp. Nonlinear nonholonomic systems, robotics, and H-infinity control theory.
Control Systems Design, 2001, pp. 1776–1781.
[19] R. Mukherjee and M. Kamon, “Almost-smooth time-invariant control
of free-flying planar space multibody systems,” IEEE Trans. Robot. Au- Masato Ishikawa (S’94–M’00) received the B.E.,
tomat., vol. 15, pp. 264–276, Apr. 1999. M.E., and Ph.D. degrees from the Tokyo Institute of
[20] G. Oriolo and M. Vendittelli, “Robust stabilization of the plate-ball ma- Technology, Tokyo, Japan, in 1994, 1996, and 2000,
nipulation system,” in Proc. IEEE Int. Conf. Robotics and Automation, respectively.
2001, pp. 91–96. From 1999 to 2002, he was a Research Associate
[21] M. Sampei, “A control strategy for a class of nonholonomic systems: at the Tokyo Institute of Technology. From 2002 to
Time-state control form and its application,” in Proc. 33th IEEE Conf. 2004, he was a Research Associate at the Department
Decision and Control, 1994, pp. 1120–1121. of Information Physics and Computing, Graduate
[22] M. Sampei, H. Kiyota, and M. Ishikawa, “Time-state control form and School of Information Science and Technology,
its application to a nonholonomic space robot,” in Proc. IFAC Symp. University of Tokyo, Tokyo, Japan. He is currently
Nonlinear Control Systems Design, 1995, pp. 759–764. an Assistant Professor with the Department of
[23] M. Sampei, S. Mizuno, M. Ishikawa, and M. Koga, “Position control of Systems Science, Graduate School of Informatics, Kyoto University, Kyoto,
a sphere between two plates,” Trans. Robot. Soc. Japan, vol. 14, no. 8, Japan. His research interests include nonlinear control theory, nonholonomic
pp. 1237–1242, 1996. mechanics, and hybrid dynamical systems and their application to robotic
[24] M. Sampei, S. Mizuno, M. Segawa, M. Ishikawa, H. Date, and D. Ya- systems.
mada, “A feedback solution to ball-plate problem based on time-state
control form,” in Proc. American Control Conf., 1999, pp. 1203–1207.
[25] H. Date, M. Sampei, D. Yamada, M. Ishikawa, and M. Koga, “Manipu-
lation problem of a ball between two parallel plates based on time-state Masanobu Koga (S’90–M’93) received the B.S,
control form,” in Proc. 38th IEEE Conf. Decision and Control, 1999, M.S., and Ph.D. degrees in control engineering from
pp. 2120–2125. the Tokyo Institute of Technology, Tokyo, Japan, in
[26] M. Sampei, H. Kiyota, M. Koga, and M. Suzuki, “Necessary and suffi- 1988, 1990, and 1993, respectively.
cient conditions for transformation of nonholonomic system into time- In 1993, he joined the faculty of the Tokyo Insti-
state control form,” in Proc. 35th IEEE Conf. Decision and Control, tute of Technology, and in 1994, moved to the Grad-
1996, pp. 4745–4746. uate School of Information Science and Engineering,
[27] H. Kiyota and M. Sampei, “An attitude control of planar nonholonomic Tokyo Institute of Technology. In 2000, he moved
space robots,” in Proc. 25th SICE Symp. Control Theory, 1996, pp. to Kyushu Institute of Technology, Fukuoka, Japan,
51–56. where he is currently an Associate Professor in the
[28] , “On the stability of a class of nonholonomic systems using time- Department of Systems Innovation and Informatics,
state control form,” in Proc. 26th SICE Symp. Control Theory, 1997, pp. Faculty of Computer Science and Systems Engineering. His current research in-
409–412. terests are in computer-aided design of control systems, system identification,
and its applications.
Dr. Koga is a Member of the Society of Instrumentation and Control Engi-
neers, and the Institute of Systems, Control and Information.

You might also like