Cobot control

Abstract

Cobots are a class of mechanically passive robotic devices,
intended for direct physical collaboration with a human operator. The
operator supplies all motive power while the cobot enforces
software-defined guiding surfaces, or constraints. Cobots are
intrinsically passive, safe devices. This is because, rather than employ
powered actuators to produce constraint forces, cobots use
“steerable” nonholonomic joints. Constraint forces are
mechanical in origin, yet software defined. The simplest possible cobot
is a unicycle which is steered by a servo system acting under computer
control, but which is moved by a human operator. The unicycle cobot
requires essentially no consideration of kinematics. Two fundamental
control modes of the unicycle cobot, “virtual caster” and
“constraint tracking”, are reviewed. More complicated
cobots, such as the three-wheeled “Scooter”, require a set
of kinematic tranformations relating configuration space to joint space.
These transformations play a role in cobot control like that of the
Jacobian in robot control

Full text

Cobot Control
Witaya Wannasuphoprasit
R. Brent Gillespie
J. Edward Colgate
Michael A. Peshkin
Department of Mechanical Engineering
Northwestern University
Evanston, IL 60208-3111
Abstract
Cobots are a class of mechanically passive robotic
devices, intended for direct physical collaboration with a
human operator. The operator supplies all motive power
while the cobot enforces software-defined guiding surfaces,
or constraints. Cobots are intrinsically passive, safe
devices. This is because, rather than employ powered
actuators to produce constraint forces, cobots use
“steerable” nonholonomic joints. Constraint forces are
mechanical in origin, yet software defined.
The simplest possible cobot is a unicycle which is
steered by a servo system acting under computer control,
but which is moved by a human operator. The unicycle
cobot requires essentially no consideration of kinematics.
Two fundamental control modes of the unicycle cobot,
“virtual caster” and “constraint tracking”, are reviewed.
More complicated cobots, such as the three-wheeled
“Scooter”, require a set of kinematic transformations
relating configuration space to joint space. These
transformations play a role in cobot control like that of
the jacobian in robot control.
1. Introduction
Co-manipulation robots, or “cobots”, are a class of
passive robotic devices which are intended for direct
collaboration with a human operator within a shared
workspace. Cobots are intended not to enhance human
strength, but to provide virtual “guiding” surfaces. The
human part of the team supplies all motive power, while
the cobot interfaces to a computer.
As an illustrative example, imagine human-cobot
collaboration in automobile assembly. Suppose that the
task is to put a large, ungainly part, such as an instrument
panel, into the vehicle body. The cobot would have two
jobs: supporting the instrument panel weight, and setting
up virtual guides which, when followed, would ensure that
no collisions could occur. The human would have the job
of pushing the instrument panel along the virtual guides
until properly positioned within the body. The human
would also be in charge of fine manipulation tasks, such
as “seating” the instrument panel.
Although the constraint forces in this example might
be quite large, a cobot would not need to generate large
actuator torques at all. This is because a cobot, unlike a
conventional robot or haptic interface, has no joint
actuators. Instead, a cobot has “steerable” nonholonomic
joints whose intrinsic mechanics provide the constraint
forces.
The essentials of cobot mechanics may be understood
by considering the simplest such device, the unicycle
cobot shown in Fig. 1. The cobot mechanism consists of
a free-rolling wheel in contact with a working surface.
The wheel’s rolling velocity is monitored by an encoder,
but it is not driven by a motor. The motor in the figure
simply steers the wheel, and cannot cause the cobot to
move. Only the operator can cause it to move, by
applying forces to the handle. A handle-mounted force
sensor monitors these user forces. The linear rails in the
figure are part of a Cartesian frame which supports the
cobot upright. These rails are also instrumented with
linear potentiometers in order to measure the global
position of the cobot.
Figure 1. The unicycle cobot.
This cobot has a two-dimensional (planar)
configuration space corresponding to all possible locations
of the unicycle assembly on the working plane. Although
the unicycle has only one degree-of-freedom, it may, by
proper steering, reach any point in the planar workspace.
Such is the nature of nonholonomic constraint. In
operation, however, virtual constraint surfaces may be
Cobot Control
Witaya Wannasuphoprasit, Brent Gillespie, J. Edward Colgate, Michael A. Peshkin
1997 International Conference on Robotics and Automation. Albuquerque, NM

defined in software to prohibit entry into excluded regions
of the plane.
The unicycle cobot has two essential modes of
operation:
a. Virtual caster mode is invoked when the cobot’s
position in its planar workspace is away from all defined
constraint surfaces. The cobot should therefore permit any
motion that the user attempts to impart. Thus, the wheel
must act something like a caster; yet, it is not a caster in
the conventional sense. Instead, it has a straight-up shaft
like a unicycle. Virtual castering arises when a handle-
mounted force sensor detects forces perpendicular to the
wheel’s rolling direction, and the wheel is steered (by a
motor) to minimize these forces. In effect, the wheel is
always steered in the direction that the user wants it to
roll.
b. Constraint tracking mode is invoked when the
user brings the cobot’s position in the plane to a place
where a constraint surface is defined. At this point, the
computer which controls the steering motor no longer
attempts to minimize force. Instead, the wheel is steered
in a direction that is tangential to the constraint surface.
The force sensor still monitors force perpendicular to the
wheel’s rolling direction. If the force would tend to push
the wheel into the constraint, it is ignored. If the force
would tend to pull the wheel off of the constraint, is
interpreted just as in the virtual caster mode. Thus, is
impossible to push the unicycle past a virtual constraint
(unless the wheel slips), but the unicycle can easily be
pulled off the constraint surface.
The unicycle cobot has some noteworthy char-
acteristics. First, although it is a one d.o.f. device (the
wheel fixes the ratio of x and y velocities), it behaves in
the virtual caster mode as though it has two d.o.f. Sec-
ond, although it uses a motor to steer, it is completely
passive in the plane of operation. Because the motor
exerts torques about an axis that passes through the
wheel/ground contact point, it does not generate any reac-
tion forces in the plane.
Cobots with larger dimensional configuration spaces
1
exist, and are discussed in [4]. The remainder of this paper
focuses on cobot control and kinematics. We begin, in
the next section, with a discussion of unicycle control.
We go on, in §3, to discuss the control of Scooter, a
tricycle cobot with nontrivial kinematics.
2. Unicycle Control
2.1 Virtual Caster
The ideal caster controller would perceptually elimi-
nate the wheel. In other words, a user manipulating the
machine would perceive it to be a point mass in the plane.
For a point mass, the acceleration and force vectors are
1
One is tempted to write “higher degree-of-freedom
cobots”, but all cobots, regardless of the configuration
space dimension, have one degree-of-freedom.
collinear and in fixed proportion. The implication for a
unicycle is that, not only must forces in the wheel
direction, F
||
, produce accelerations of a
||
= F
||
/M, but
forces normal to the wheel, F
⊥
, must similarly produce
accelerations of a
⊥
= F
⊥
/M. A very simple kinematic
analysis, however, shows that a wheel traveling at a speed
u with a steering angular velocity ω, has an instantaneous
normal acceleration of a
⊥
= uω. Thus, we can obtain a
prescription for the steering velocity in response to user
forces which would result in point-mass-like behavior:
ω =
F
⊥
uM
(1)
Note that u in Eq. 1 is not under our control but is
determined by the user. This equation indicates that the
problem of virtual caster control is fundamentally non-
linear: the correct sign of the steering velocity is
determined by the product of the signs of F
⊥
and u, which
cannot be approximated by a linear relation. Eq. 1 also
indicates that, for a given normal force, the steering
velocity scales inversely with the translational velocity.
Because of this, there is a singularity at zero speed. At
zero speed, it is not physically possible to make the
unicycle behave like a particle.
An in-depth discussion of unicycle caster control,
including implementation issues (e.g., handling the zero
velocity singularity; finite sensor resolution) can be found
in [2].
2.2 Path Tracking
In this paper, we will consider only the case of
bilateral constraint (path tracking) rather than unilateral
constraint (virtual wall). For the unicycle cobot, a
bilateral constraint is easily made unilateral with a
software switch [1, 2]. We will discuss a nominal, or
feedforward, controller first, followed by feedback control.
A standing assumption is that the path to be tracked,
R
o
(s), is parameterized in terms of its own path length, s.
2.2.1 Feedforward Control
An appropriate feedforward controller can be derived
by assuming perfect path tracking. In such a case, the
unicycle would at all times be tangent to the path, as
illustrated in Fig. 2. Moreover, the speed, u, of the
command
path
T
κN
Figure 2. Top view of a unicycle cobot tracking a path

unicycle along the path, and the curvature, κ, of the path
would together determine the necessary steering velocity:
ω = κu (2)
Unfortunately, a number of non-idealities including
steering dynamics and sideslip, will conspire to ensure that
this feedforward controller alone will not keep the unicycle
on an intended path. Thus, it is necessary to consider
feedback corrections.
2.2.2 Feedback Control
Fig. 3 is an illustration of a unicycle which is off of
its intended path. In such a case, the C-space
configuration of the unicycle cobot (R = [x,y]
T
) is not
what the controller is expecting (R
o
(s), where s is the
measured path length). Moreover, the cobot heading
2
(T)
is, in general, not parallel to the expected heading (T
o
).
The control policy that corrects for these errors has two
components:
1. s is replaced with s’, where R
o
(s’) is the closest point
on the command path to the cobot’s actual
configuration (R).
2. κ in Eq. 2 is replaced with κ + δκ, where δκ is
determined on the basis of the configuration and
heading errors, and is intended to steer the unicycle
back toward the command path.
κN
T
R
T
o
(s)
command
path
R
o
(s)
actual path
Figure 3. Imperfect path tracking
The closest point computation required to find s’ will, in
general, be complicated to perform. However, if we make
the reasonable assumption that deviations from the path
are small (i.e., our path tracking controller works well),
then it is evident (see Fig. 3) that the path length error (s’
- s) is simply the length of (R - R
o
(s)) projected onto T
o
.
Thus, we define:
s’ = s + (R - R
o
(s)) ⋅ T
o
(s) (3)
When the actual cobot configuration is compared to R
o
(s’)
there are of course still errors. These errors are of two
types:
• Displacement — ∆R = R - R
o
(s’) is a vector
2
The “heading” of a cobot is represented by a unit tangent
vector in C-space.
approximately normal to the command path
representing the distance that the cobot is off the path.
• Heading — ∆T = T - T
o
(s’) is a vector approximately
normal to the command path representing the error in
cobot heading.
A standing assumption is that the cobot to be controlled is
outfitted with the sensors necessary to measure these
errors. In the case of the cobot pictured in Fig. 1, linear
potentiometers provide C-space configuration (R) and a
steering shaft encoder provides heading (T).
An intuitive approach to feedback control is simply to
adjust the steering velocity (ω) in order to compensate for
both types of errors. In fact, it is somewhat more
powerful to think in terms of adjusting a curvature
command (κ), because the path velocity u is solely at the
discretion of the operator. Thus, we would expect that a
reasonable form for the controller would be:
ω = u




κ
o
(s’) - N
o
(s’)⋅




G
1
L
2
∆R +
G
2
L
∆T (4)
κ
o
(s’) is the feedforward term from §2.2.1. G
1
and G
2
are
control gains, and L is a scale factor used to ensure
consistent units. L may also be interpreted as a lookahead
distance [2, 3].
Gain selection as well as the stability of this type of
controller have been discussed in depth in [2]. Further
discussion of these issues will be omitted here, due to
space limitations. Fig. 4, however, shows recorded data in
both caster and constraint modes for the unicycle cobot.
2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
x position (inches)
y position (inches)

3. Control of Scooter
3.1 Scooter
Scooter, a redundant tricycle cobot, is pictured in Fig.
5. Scooter has been built primarily as a testbed for
exploring the kinematics and control of higher
dimensional cobots.
Figure 5. Scooter
The configuration space of Scooter is that of a planar
rigid body (R = [x, y, lθ]
T
). Only two wheels are needed
to produce one degree-of-freedom motion in this space;
however, Scooter is outfitted with a third wheel to
eliminate the need for external support, and to eliminate a
singularity that occurs with only two wheels [1].
It is our assumption that, in controlling Scooter or
other high dimensional cobots, it will be desirable to plan
paths and constraint surfaces in C-space, rather than in the
space of individual joints. Thus, the first topic we address
here is kinematic transformations from C-space to joint
space.
3.2 Kinematics
In the case of Scooter, the C-space configuration may
be described as the position (x, y) and orientation (θ) of
the handle location (in the middle of the equilateral
triangle), measured relative to a known starting location in
a global frame. The orientation, however, is generally
scaled by a characteristic length measure, l, to minimize
numerical difficulties. Thus, R = [x, y, lθ]
T
.
The path followed by wheel i can easily be derived
from a knowledge of R and of the vector that runs from
the handle location to the wheel steering axis. The
operations involved are simply translation and rotation.
Thus, it is possible to write:
r
i
=




x
i
y
i
=




L
ix
(R)
L
iy
(R)
= L
i
(R), i = 1,2,3 (5)
Eq. 5 indicates that, unlike the unicycle cobot,
Scooter exhibits non-trivial kinematics relating its C-
space path to its wheel paths. Like the unicycle, however,
it is necessary to steer individual wheels in order to control
Scooter. This suggests that several C-space-to-joint-space
kinematic transformations are necessary (one for each
wheel). In the following, we develop the transformation
relations necessary to relate the C-space path tangent to
the joint (wheel) path tangents, and the C-space path
curvature to the joint path curvatures.
3.2.1 Tangent Transformation
C-space and joint space unit tangent vectors are
defined as follows (note that “joint space” here refers to the
space associated with a single wheel):
T =
dR
ds
; t
i
=
dr
i
ds
i
, i = 1,2 3 (6)
Here, s
i
is the arc length along the path traced out by
wheel i. A relation between these two types of tangents
can be obtained by differentiating Eq. 5 with respect to s
i
:
t
i
=
dr
i
ds
i
=
∂L
i
∂R
dR
ds
ds
ds
i
(7)
The first term on the right hand side can be recognized as
an ordinary 2×3 Jacobian, which we will name J
i
, from
cobot C-space to the two-dimensional path space of wheel
i. The second term is the C-space tangent, and the third
term is a local path length expansion/contraction ratio.
This ratio is a scalar, and its magnitude is necessarily that
which makes t
i
a unit vector. Thus:
t
i
=
J
i
T
|J
i
T|
(8)
3.2.2 Curvature Transformation
C-space and joint space curvatures (κ, κ
i
) and unit
normal vectors (N, n
i
) are defined as follows:
κN =
dT
ds
; κ
i
n
i
=
dt
i
ds
i
, i = 1,2 3 (9)
An expression for κ
i
n
i
can be obtained by differentiating
Eq. 8 with respect to s
i
. Because Eq. 8 involves a quotient
of terms, all of which depend upon s, the operations are
fairly tedious, though straightforward. Details will be
omitted. The result is:
κ
i
n
i
=
I - t
i
t
i
T
|J
i
T|
2








T
T
H
ix
T
T
T
H
iy
T
+ J
i
κN
(10)
H
ix
and H
iy
are 3×3 matrices defined as:
H
ix
=
∂
2
L
ix
∂R
2
; H
iy
=
∂
2
L
iy
∂R
2
(11)
They may be thought of as representing the spatial rate of
change of the Jacobian, J
i
. This effect, even in the
absence of curvature in C-space, can result in joint space
curvature (i.e., curvature of the path that a wheel traces out
on the ground). The term I - t
i
t
i
T
may be recognized as a

projection matrix, which ensures that n
i
will be normal to
t
i
.
3.3 Virtual Caster
Unlike the unicycle cobot, the apparent dof of Scooter
may vary from 1 to 3 (if the redundant wheel is
misaligned, even zero dof is a possibility). Although
quite interesting and important in its own right, the
mathematics involved in setting up 2 dof behaviors, and
more importantly, switching between 1, 2 and 3 dof
behaviors, goes beyond the scope of this paper. Here, we
will restrict attention to a fully castered, 3 dof mode. In
the next section, we will focus on 1 dof path tracking.
C-space caster control for Scooter is much like caster
control for the unicycle cobot. In addition to
instantaneous configuration (which figures into Jacobians,
etc.), two key measurements are necessary: velocity (both
magnitude, u, and direction, T), and operator-applied force
(F = [F
x
, F
y
, l
-1
τ]
T
).
In order to make the cobot behave as an unconstrained
rigid body, the force measurement can be used to generate
a desired acceleration. For instance, this acceleration
might be A = [F
x
/m, F
y
/m, τl/I]
T
, where m and I are
estimates of cobot mass and moment, respectively. These
estimates need not be terribly accurate. A more general
interpretation of them is as caster controller gains.
Moreover, this is only one prescription for generating a
desired acceleration vector. More general ones involving
non-diagonal inertia matrices can be readily imagined.
Once a desired acceleration, A, is available, caster
control is straightforward. The idea is to use A, T and u
to predict a C-space path. κN for this path is found and
inserted, along with T, into Eq. 10 to produce the
necessary joint space curvatures, from which wheel
steering velocities are determined. The following relations
between the time and path derivatives of R are useful:
dR
dt
= Tu (12)
d
2
R
dt
2
= A = κNu
2
+ T˙u
(13)
Also useful is the following result, which may be found
by straightforward differentiation of u:
˙u
= T
T
A (14)
Combining Eqs. 13 and 14, the C-space curvature vector
is found:
κN =
1
u
2
[I - TT
T
] A (15)
This result is pleasingly simple: the component of the
desired acceleration which is normal to the C-space path is
selected by the projection matrix and scaled by u
2
to
produce the curvature. Eq. 10 can then be used to compute
the joint space curvatures.
The steering velocity commands are each the product
of a curvature and a speed. The applicable speed is, of
course, the path speed at the joint, u
i
. It is readily shown
that:
t
i
u
i
= J
i
Tu i = 1,2,3 (16)
As a final step, it must be recognized that, in the case
of Scooter, the steering motors are mounted in the moving
frame of Scooter rather than the ground frame. Thus, the
steering velocity commands must be adjusted for the
rotational velocity of Scooter:
ω
i
=


J
i
Tu ×
I - t
i
t
i
T
|J
i
T|
2


I - t
i
t
i
T
|J
i
T|
2








T
T
H
ix
T
T
T
H
iy
T
+ J
i
1
u
2
[I - TT
T
] A
-
˙
θ
i = 1,2,3 (17)
3.4 Path Tracking
In the above discussion, we saw that C-space caster
control of Scooter is in large part analogous to caster
control of the unicycle cobot: a desired normal
acceleration is determined, and used to command steering
velocities. It will be shown in this section that the
analogy is equally strong in the case of constraint
tracking. Moreover, we are motivated to investigate
constraint tracking in C-space by the belief that path
planning will be more naturally accomplished in C-space
than in joint space. As with the unicycle, we will break
the discussion into feedforward and feedback control.
3.4.1 Feedforward Control
Feedforward control is developed on the basis of ideal
path tracking. In this case, the cobot heading, T, and
curvature vector, κN, are assumed to be identical to those
of the command path (T
o
and κ
o
N
o
), and the cobot is
assumed to be on the path. As before, the command path
is assumed to be parameterized in terms of its own path
length. The feedforward command for each joint can then
be determined by using Eqs. 10 and 16, and following the
procedure decribed in the previous section.
3.4.2 Feedback Control
The control policy that corrects for configuration and
heading errors has two components. These components
are analogous to those for the unicycle cobot, however,
rather than make adjustments to the scalar curvature, it
becomes necessary to make adjustments to the curvature
vector (i.e., both κ and N):
1. s is replaced with s’, where R
o
(s’) is the closest point
on the command path to the cobot’s actual
configuration (R).
2. κN is replaced with κ
o
(s’)N
o
(s’) + δ(κN), where
δ(κN) is determined on the basis of the configuration
and heading errors, and is intended to steer Scooter
back toward the command path in C-space.
s’ continues to be specified according to Eq. 3, and when
the cobot configuration is compared to R
o
(s’) there
continue to be two types of errors: displacement, ∆R, and
heading, ∆T. These vectors, while approximately normal

to the path, will generally not be parallel to N
o
(s’). Thus,
the correction to be made will not be simply a scaling of
the curvature vector, but also a redirecting of this vector.
A reasonable form for δ(κN) would be similar to the
curvature correction employed for unicycle control (Eq. 4):
δ(κN) = -
G
1
L
2
∆R -
G
2
L
∆T (18)
Somewhat greater insight into Eq. 4 can be obtained by
considering the terms individually. Let’s begin with the
displacement term. Associated with it is a circle that is
tangent to T and lies in the plane common to T and ∆R.
The magnitude of this term also determines the circle’s
curvature. These points are illustrated in Fig. 6. Note
that N
o
(s’) does not generally lie in the plane of the circle.
This circle may be considered the instantaneous
“return path” associated with the displacement term.
Indeed, the circle will intersect the command path tangent,
T
o
(s’), at a “return distance” of L(2/G
1
)
1/2
.
T
o
(s')
T
"return
distance"
command path
R
o
(s') - R
L
2
G
1
R
o
(s') - R
circle of radius:
x
y
l
θ
Figure 6. Geometry of circular “return path” induced by
displacement term of feedback controller.
There is a similar circle associated with the heading
term, and it produces a “return distance” of 2L/G
2
. If we
make the gain selection G
1
= G
2
= 2, then it is appropriate
to speak of L as a “lookahead distance”, i.e., a
characteristic C-space path length for error recovery.
Returning now to the design of the constraint tracking
controller, it is necessary only to introduce the curvature
correction (Eq. 18) into the transformation relation (Eq.
10), and follow the procedure developed for caster control.
The resulting control law is:
ω
i
=


J
i
Tu ×
I - t
i
t
i
T
|J
i
T|
2


I - t
i
t
i
T
|J
i
T|
2








T
T
H
ix
T
T
T
H
iy
T
+ J
i
(κ
o
(s’)N
o
(s’) + δ(κN))
-
˙
θ
i = 1,2,3 (19)
A discussion of gain selection and stability will be
deferred to future publications. These matters, as well as
an experimental investigation involving Scooter, are the
subject of current work.
4. Conclusions
The fundamentals of virtual caster and constraint
tracking control of cobots have been introduced. A key
concept is that of a path in cobot C-space parameterized by
path length, and having tangent and curvature vectors. In
addition, tangent and curvature transformations between C-
space and joint space are seen to play a key role. The
joint space of a cobot is unlike that of a conventional
robot: associated with each cobot joint is a two-
dimensional path (e.g., the path that a wheel traces out on
the ground).
It is interesting to note that most of the ideas that
have been introduced here apply to cobots of higher
dimensional C-space, and even cobots which employ
steerable nonholonomic joints other than wheels (such as
those discussed in [4]).
For more information about cobots, the reader is
invited to visit http://www.ece.nwu.edu/~peshkin/cobot/.
Acknowledgements
The authors gratefully acknowledge the support of the GM
Foundation and the National Science Foundation. We are
particularly grateful for the intellectual input of Prasad
Akella, and all the members of the cobot group at
Northwestern.
References
[1] Colgate, J. E., M. A. Peshkin and W.
Wannasuphoprasit. Nonholonomic Haptic Display.
IEEE International Conference on Robotics and
Automation. Minneapolis. pp. 539-544, 1996.
[2] Colgate, J. E., W. Wannasuphoprasit and M. A.
Peshkin. Cobots: Robots for Collaboration with
Human Operators. International Mechanical
Engineering Congress and Exposition. Atlanta. pp.
433-440, ASME, 1996.
[3] Ollero, A. and G. Heredia. Stability Analysis of
Mobile Robot Path Tracking. IEEE/RSJ International
Conference on Intelligent Robots and Systems
(IROS). Pittsburgh, PA. pp. 461-466, 1995.
[4] Peshkin, M., J. E. Colgate and C. Moore. Constraint
Machines Based on Continuously Variable
Transmissions, for Haptic Interaction with People.
IEEE International Conference on Robotics and
Automation. pp. 551-556, 1996.