Impedance Control: An Approach

to Manipulation:
Neville Mogan Part SI—Implementation
Associate Professor,
Department of Mechanical Engineering This three-part paper presents an approach to the control of dynamic interaction
and Laboratory for Manufacturing between a manipulator and its environment. Part I presented the theoretical
and Productivity,
reasoning behind impedance control. In Part II the implementation of impedance
Massachusetts Institute of Technology,
Cambridge, Mass. 02139 control is considered. A feedback control algorithm for imposing a desired car-
tesian impedance on the end-point of a nonlinear manipulator is presented. This
algorithm completely eliminates the need to solve the "inverse kinematics problem"
in robot motion control. The modulation of end-point impedance without using
feedback control is also considered, and it is shown that apparently "redundant"
actuators and degrees of freedom such vs exist in the primate musculoskeletal
system may be used to modulate end-point impedance and may play an essential
functional role in the control of dynamic interaction.

Introduction
Most successful applications of industrial robots to date environment, to ensure physical compatibility with the en-
have been based on position control, in which the robot is vironmental admittance, something has to give, and the
treated essentially as an isolated system. However, many manipulator should assume the behavior of an impedance.
practical tasks to be performed by an industrial robot or an Thus a general strategy for controlling a manipulator is to
amputee with a prosthesis fundamentally require dynamic control its motion (as in conventional robot control) and in
interaction. The work presented in this three-part paper is an addition give it a "disturbance response" for deviations from
attempt to define a unified approach to manipulation which is that motion which has the form of an impedance. The
sufficiently general to control manipulation under these dynamic interaction between manipulator and environment
circumstances. may then be modulated, regulated, and controlled by
In Part I this approach was developed by starting with the changing that impedance.
reasonable postulate that no controller can make the This second part of the paper presents some techniques for
manipulator appear to the environment as anything other controlling the impedance of a general nonlinear multiaxis
than a physical system. An important consequence of manipulator.
dynamic interaction between two physical systems such as a
manipulator and its environment is that one must physically Implementation of Impedance Control
complement the other: Along any degree of freedom, if one is
an impedance, the other must be an admittance and vice A distinction between impedance control and the more
versa. conventional approaches to manipulator control is that the
One of the difficulties of controlling manipulation stems controller attempts to implement a dynamic relation between
from the fact that while the bulk of existing control theory manipulator variables such as end-point position and force
applies to linear systems, manipulation is a fundamentally rather than just control these variables alone. This change in
nonlinear problem. The familiar concepts of impedance and perspective results in a simplification of several control
admittance are usually applied to linear systems and regarded problems.
as equivalent and interchangeable. As shown in Part I, for a Most of our work to date [3, 6, 13, 14, 16] has focused on
nonlinear system, the distinction between the two is fun- controlling the impedance of a manipulator as seen at its
damental. "port of interaction" with the environment, its end effector.
Now, for almost all manipulatory tasks the environment at A substantial body of literature has been published on
least contains inertias and kinematic constraints, physical methods for implementing a planned end effector cartesian
systems which accept force inputs and which determine their path [5, 27, 28, 32, 34, 35]. The approach is widely used in the
motion in response and are properly described as admittances. control of industrial manipulators and there is some evidence
When a manipulator is mechanically coupled to such an of a comparable strategy of motion control in biological
systems [1, 24]. Following the lead from this prior work we
Contributed by the Dynamic Systems and Control Division for publication in
have investigated ways of presenting the environment with a
the JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL. Manuscript dynamic behavior which is simple when expressed in
received by the Dynamic Systems and Control Division, June 1983. workspace (e.g., Cartesian) coordinates.

8 / V o l . 107, MARCH 1985 Transactions of the ASME

Copyright © 1985 by ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 The lowest-order term in any impedance is the static the manipulator need be computed. This is a direct con-
relation between output force and input displacement, a sequence of the care which was taken to ensure that the
stiffness. If, in common with much of the current work on desired behavior was compatible with the fundamental
robot control, we assume actuators capable of generating mechanics of manipulation and was expressed as an im-
commanded forces (or torques), Tact, sensors capable of pedance.
observing actuator position (or angle), 6, and a purely Another important term in the manipulator impedance is
kinematic relation (i.e., no structural elastic effects) between the relation between force and velocity. Again, given the
actuator position and end-point position1, X = L(6), it is above assumptions, it is straightforward to define a feedback
straightforward to design a feedback control law to im- law to implement in actuator coordinates a desired relation
plement in actuator coordinates a desired relation between between end-point force and end-point velocity such as:
end-point (interface) force, Fint, and position, X. Defining
the desired equilibrium position for the end-point in the Fint = fi[V0-V] (5)
absence of environmental forces (the virtual position) as X0, a From the manipulator kinematics:
general form for the desired force-position relation is: V = J(0)a> (6)
Fint=ATXo-X]- (1) The required relation in actuator coordinates is:
Compute the Jacobian, J(0): Tact = J'(0)fl[Vo - J(0)co] (7)
dX = 3(6)d6 (2) Again note that the relation B[\0 - V] need not be linear and
From the principal of virtual work: that inversion of the Jacobian is not required.
The dynamic behavior to be imposed on the manipulator
Tact = J'(0)Fint (3) should be as simple as possible, but no simpler. The foregoing
The required relation in actuator coordinates is: equations take no account of the inertial, frictional, or
Tact = J'(0)A-[Xo-L(0)] (4) gravitational dynamics of the manipulator. Under some
circumstances this may be reasonable, but in many situations
No restriction of linearity has been placed on the relation these effects cannot be neglected. To ensure dynamic
K[X0 - X ] , and the displacement of X from X0 need not be feasibility the choice of the impedance to be imposed should
small. Note that in this equation the inverse Jacobian is not be based on the dominant dynamic behavior of the
required. manipulator. The choice is a tradeoff between keeping the
complexity of the controller within manageable limits while
Inverting the kinematic equations of a manipulator to ensuring that imposed behavior adequately reflects the real
determine the time-history of actuator (joint) positions dynamic behavior of the controlled system. As a result it
required to produce a desired time-history of end-point depends both on the manipulator itself and on the en-
positions has been described as one of the most difficult vironment in which it operates. For example, a manipulator
problems in robot control [28]. For some manipulators (e.g., intended for underwater applications will operate in a
those with nonintersecting wrist joint axes) no explicit (closed- predominantly viscous environment and it may be reasonable
form) algebraic solution may be possible. However, if K[X0 - to ignore inertial effects. In contrast, a manipulator intended
X] is chosen so as to make the end-point sufficiently for operation in a free-fall orbit will encounter a
stiff, then a controller which implements equation (4) will predominantly inertial environment. For terrestrial ap-
accomplish Cartesian end-point position control and the need plications (which have been the main concern of our work)
to solve the "inverse kinematics problem" has been com- both gravitational and inertial effects are important, and the
pletely eliminated. Only the forward kinematic equations for dominant dynamic behavior is that of a mass driven by
'Throughout this paper, "position" will refer to both location and orien-
motion-dependent forces, second order in displacement along
tation, and "force" will refer to both force and moment. each degree of freedom.

Y admittance h = generalized momentum in

Z = impedance actuator coordinates
modulation by command M = inertia tensor in end-point P = generalized momentum in
•Ac]
set coordinates end-point coordinates
m = mass H(.) = Hamiltonian
Sf _ flow source
Se = effort source I = inertia T,Ti,T2 = torque
Fext _ external force t = time #1.02 = absolute j oint angle
F(-) = noninertial impedance Pi 1P2 = relative joint angle
Fint = interface force Me = environmental inertia -^1 J - ^ 2 = link lengths
X = end-point position tensor Ke = net stiffness due to elbow
Xo = commanded (virtual) 1(6) = inertia tensor in actuator muscles
position coordinates Ks = net stiffness due to
V = end-point velocity C(0,co) = inertial coupling torques shoulder muscles
V0 = commanded (virtual) K(co) = velocity-dependent torques Kt = net stiffness due to two-
velocity S(6) = position-dependent joint muscles
K[.] = force/displacement rela- torques Kx = stiffness tensor in end-
tion G(0,«) = accelerative coupling terms point coordinates
£[•] = force/velocity relation Tact = actuator force or torque Ko = stiffness tensor in joint
6 = actuator position or angle Tint = interface torques coordinates
CO = actuator velocity Y(6) = mobility tensor in actuator Ep = potential energy
U-) = linkage kinematic equa- coordinates Ek = kinetic energy
tions W(6) = mobility tensor in end- Ek* = kinetic coenergy
3(6) = Jacobian point coordinates X,X2 = eigenvalues

Journal of Dynamic Systems, Measurement, and Control MARCH 1985, Vol. 107/9

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 BONOS REPRESENT VECTOR QUANTITIES All of the parameters in this expression are implicitly assumed
I to be functions of the set of control commands (c) and of
F„1'S, = ^ | 1 y -. 0 _] S r V 0 ^ j c ) time.
This set of assumptions defines a target behavior which
I 1 includes inertial effects. The first two terms are the position-
iL ' \ and velocity-dependent impedances of equations (1) and (5).
EXTERNAL Y ' Z • fc) If the environment is a simple rigid body acted on by un-
SOURCE I ' ' predictable (or merely unpredicted) forces, its dynamic
I
I equations are:
ENVIRONMENTAL | CONTROLLED MANIPULATOR MedV/dt = Fext + ¥mt (12)
ADMITTANCE |
Fig. 1 A bond graph equivalent network representation of an im- and the coupled equations of motion for the complete system
pedance-controlled manipulator interacting with an environmental of figure 1 are:
admittance. Each bond represents a vector of power flows along
multiple degrees of freedom. (Me+M)dV/dt=K[X0-X]+B[\0 - V] +Fext (13)
When the manipulator is decoupled from its environment Note that in this case both the coupled and uncoupled
the term due to the environmental admittance disappears, and equations for the system have the same second-order form.
in principle the manipulator alone need exhibit no mass-like To implement the target behavior of equation (11), one
behavior. In practice, the uncoupled manipulator still has approach we have used is to express the desired Cartesian
inertia (albeit nonlinear and configuration-dependent). This coordinate impedance in actuator coordinates (the kinematic
means that the controlled system, both with the manipulator transformations between actuator coordinates and end-point
coupled to and uncoupled from its environment, can be coordinates provides sufficient information to do this) and
represented by an admittance coupled to an impedance as then use a model of the manipulator dynamics to derive the
shown in Fig. 1. required controller equations. Assuming that the kinematic,
No physically realisable strategy can eliminate the inertial inertial, gravitational, and frictional effects provide an
effects of a manipulator but the apparent inertia seen at the adequate model of the manipulator dynamics as follows:
end effector can be modified. The approach we have taken to I(8)dco/dt + C(6,a>)+V(co) + S(d) = Tact+Tint (14)
deal with inertial manipulator behavior is to "mask" the true
nonlinear inertial dynamics of the manipulator and impose an expression for the required actuator torque as a function of
simpler dynamics, those of a rigid body. Most manipulatory actuator position and velocity and end-point force can be
tasks are fundamentally described in relative coordinates, that derived by straightforward substitution (see Appendix I):
is, in terms of displacements and rotations with respect to a Tact = 1(6)3 ~,(6)M^lK[X0-L(6)]+S(6)
workpiece, tool or fixture whose location in the workspace is
not known in advance with certainty. As a result, task + 1(6)5',(d)M-iB[\0-3(d)u] + V(o)
planning and execution will be simplified if the end-point + I(ff)J~1(e)M-iFmt-Jt(9)Fmt
inertial behavior is modified to be that of a rigid body with an
inertia tensor which remains invariant under translation and - 1(6)3 ~1(6)G(6,u) + C(6,w) (15)
rotation of the coordinate axes. This is achieved if: This equation expresses the required behavior to be
provided by the controller as a nonlinear impedance in ac-
m 0 0 0 0 0 tuator coordinates. It may be viewed as a nonlinear feedback
law relating actuator torques to observations of actuator
0 m 0 0 0 0
position, velocity and interface force. The input (command)
0 0 m 0 0 0 variables are the desired cartesian position (and velocity) and
the terms of the desired (possibly nonlinear) cartesian
0 0 0 / 0 0 dynamic behavior characterized by M, /?[•] and AT*].
The feasibility of this approach to cartesian impedance
0 0 0 0 / 0 control has been investigated [6,16] by implementing this
nonlinear control law to impose cartesian end-point dynamics
0 0 0 0 0 / on a servo-controlled, planar, two-link mechanism (similar to
the nonlinear linkage in a SCARA 2 robot). A simple analysis
This is the inertia tensor of a rigid body such as a cube of estimating the computation required to implement this
uniform density. This inertia tensor eliminates the angular controller on a six-degree-of-freedom manipulator indicated
velocity product terms in the Euler equations for the motion that the computational burden is comparable to "exact"
of a rigid body [8] and ensures that if the system is at rest the approaches to generating forward-path manipulator com-
applied force and the resulting acceleration vectors are mands such as the recursive LaGrangian [17] and Newton-
colinear. Euler [21] methods or the configuration space method [18].
To represent the dominant second-order behavior of the If the interface forces and torques in equations (11) and (15)
manipulator the noninertial interface forces are assumed to are eliminated and the position- and velocity-dependent terms
depend only on displacement, velocity and time: reduced to linear diagonal forms, this implementation of
Fint = F(X,V)-MdV/dt (9) impedance control resembles the resolved acceleration
method [22]. However, unlike the resolved acceleration
If the noninertial behavior to be imposed is nodic, it may be
method, the impedance control algorithm presented above is
written in terms of a displacement from a commanded (time-
based on desired end-point behaviour which may be chosen
varying) position X 0 : rationally using approaches such as the optimisation
Fint = F(X0-X,Y0-\)-MdV/dt (10) technique presented in Part III. Furthermore, the impedance
control algorithm includes terms for coping with external
Although there may be cases in which coupled nonlinear "disturbances." Without the external "disturbance" terms
viscoelastic behavior is useful, for simplicity the position- and (which have no counterpart in the resolved acceleration
velocity-dependent terms may be separated: algorithm) the manipulator is not capable of. controlled
Fint = K[X0-X]+B[V0-V]~MdY/dt (11) 2
Selective Compliance Assembly Robot Arm [23].

10/Vol. 107, MARCH 1985 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 mechanical interaction with its environment. Note also that
the above approach to defining the controller equations is not
restricted to commanded linear behavior and can be applied
equally well to achieve the more general coupled nonlinear
behavior of equation (9).
It is not claimed that the above algorithm is the only way to (a)
achieve a desired end-point impedance. It is presented here
only to demonstrate that a control law capable of modulating
the end-point impedance of a manipulator may be for-
mulated. The controller of equation (15) was designed by a
technique which is similar to pole-placement methods [31] in
that the desired behaviour and a model of the actual

3
behaviour of the manipulator were compared algebraically to
derive the controller equations. In common with most ap-
proaches to manipulator control the approach is based on a
model which ignored many aspects of real manipulator
performance, particularly the dynamics of the actuators and
the transmission system. Furthermore, like many other ap-
proaches the method assumes that the Jacobian is invertible. (b) (c) (d)
This technique is, of course, only one possible approach to Fig. 2 A schematic representation of the influence of kinematic
the design of a controller for implementing a desired cartesian redundancies on the mobility (inverse effective mass) of the end-point
impedance, and, if one may draw from linear systems design of a planar linkage. The ellipsoid of gyration associated with the
experience without overstretching the analogy to pole- mobility tensor is shown in (a). The eigenvalues of the mobility tensor
placements methods, it is not even likely to be the best. Other are inversely proportional to the effective mass in the direction of the
corresponding eigenvectors and the square root of their ratio deter-
approaches to controller design such as the model-referenced mines the ratio of the major and minor axes of the ellipsoid, which are
adaptive control method [9] will probably be useful. colinear with the eigenvectors. For a planar, three-member linkage with
links of uniform density and cross section and lengths in the ratio 1: 2:
3 the effect on the ellipsoid of gyration of changing the linkage con-
Impedance Modulation Without Feedback figuration for a fixed position of the end-point is shown in (b), (c), and
<d>.
Modulation of end-point impedance using feedback
strategies is not the only way to control the dynamic behavior suggested. The elements of the mobility tensor in general will
of a manipulator, nor is it always the best. This is particularly depend on the manipulator configuration.
evident in a biological system. One of the most distinctive At any given configuration, the kinematic transformations
features of the primate neural control system is the between joint angles and end-point coordinates define not
unavoidable delay associated with neural transmission. The only the relations between generalized displacements, flows
shortest time for information to get from peripheral sensors and efforts in the two coordinate frames, (see equations (2),
(e.g., in the muscles or skin) in the human arm to the higher (3), and (6)) they also define the relation between the
levels of the central nervous system (e.g., the cortex) and back generalised momenta in joint coordinates, h, and end-point
to the actuators of the arm is 70 milliseconds, and loop coordinates, p, through the Jacobian (see Appendix II):
transmission delays of 100 to 150 milliseconds are typical [29].
This problem is further exacerbated if significant com- h = J'(0)p (17)
putation is required (the response time to a visual stimulus is Consequently, the mobility tensor in end-point coordinates
somewhere between 200 and 250 milliseconds). The ef- W{6) is related to the mobility in joint coordinates Y{&) as
fectiveness of feedback control in the presence of a delay of follows:
this magnitude is severely limited, particularly in dealing with
tasks involving dynamic interaction. Yet primates excel at (18)
controlling dynamic interactions; How do they do that? W(0) = 3(0)Y{0)3<{8) (19)
One alternative to feedback which we have explored is the The physical meaning of the end-point mobility tensor is that
use of redundancies: "excess" actuators or "excess" skeletal if the system is at rest (zero velocity) then a force vector
degrees of freedom. From a purely kinematic standpoint the applied to the end-point causes an acceleration vector (not
neuromuscular system is multiply redundant. For example, necessarily co-linear with the applied force) which is obtained
the kinematic chain connecting the wrist joint to the chest by premultiplying the force vector by the mobility tensor (see
(clavicle, scapula, humerus, radius and ulna) has considerably Appendix II).
more degrees of freedom than those required to specify the Note that the Jacobian in the above equation need not be
position (and orientation) of the hand in cartesian coor- square, and that the end-point mobility is configuration
dinates. These skeletal redundancies can serve to provide a dependent. As a result, redundant degrees of freedom can be
measure of control over the inertial component of the end- used to modulate the end-point mobility. Consider the
point dynamics. simplified three-link model of the primate upper extremity
In considering the apparent inertial behaviour of the end- (arm, forearm and hand, each considered to be rigid bodies,
point it is useful to remember that an inertia is fundamentally linked by simple pin-joints) moving in a plane as shown in
an admittance; flow (velocity) is determined as a response to Fig. 2. For simplicity, assume the links are rods of uniform
impressed effort (force). Dealing with kinematic redundancy density with lengths in the ratio of 1: 2: 3.
is considerably simplified if the constitutive equations are Any real linkage such as the skeleton is a generalised kinetic
written as a relation determining generalised velocity, a>, (e.g., energy storage system. Kinetic energy is always a quadratic
the velocities of the manipulator joints) as a function of form in momentum:
generalized momentum, h:
Ek=lAWY(6)\\ (20)
o>=Y(6)h (16)
Thus the locus of deviations of the generalised momentum
Y{6) is the inverse of the more commonly used inertia tensor, from zero for which the kinetic energy is constant is an
and to help distinguish the two, the term "mobility" is ellipsoid, the "ellipsoid of gyration" [33]. It graphically

Journal of Dynamic Systems, Measurement, and Control MARCH 1985, Vol. 107/11

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Table 1 Variation of apparent end-point mass with linkage
configuration
Distal link Effective mass Effective mass
orientation X\ -direction X2 -direction
(degrees) (kgm) (kgm)
90 0.322 1.823
135 0.568 0.568
180 1.824 0.323
Link Lengths: 1, 2, 3 meters; Linear density: I kgm/m

represents the directional properties of the mobility tensor.

The eigenvalues of the symmetric mobility tensor define the
size and shape and the eigenvectors the orientation of the
ellipsoid of gyration (see Appendix II). An ellipsoid of
gyration can be associated with the mobility tensor in any
coordinate frame, e.g., end-point coordinates (see Fig. 2(a)). (b)
Figures 2(b) through 2(d) show the profound effect on the
ellipsoid of gyration of changes in arm configuration while WORKSPACE
keeping the position of the end-point fixed. The inertial
resistance to a force applied radially inward toward the
shoulder (vertically downward in the figure) changes by
almost a factor of six as the hand rotates through ninety
degrees (see Table 1). In the configuration of Fig. 2(d) the
applied force has to accelerate all three links; in that of Fig.
2(b) it primarily has to accelerate the distal link. Clearly,
kinematic redundancies in a linkage provide a vehicle for
Fig. 3 A schematic representation of the influence of the polyarticular
changing the way the end-point will react to external muscles of the primate upper extremity on the range of end-point
disturbances without recourse to feedback strategies. stiffnesses which may be achieved without recourse to feedback
As an aside, an alternative representation of inertial strategies by simultaneous activation of opposing muscles. The
behavior is via the ellipsoid of inertia [33]. Asada [4] has ellipsoid associated with the symmetric differential stiffness tensor is
shown in (a). The eigenvalues of the stiffness tensor are proportional to
suggested its use as a tool for designing robot mechanisms. the stiffness in the direction of the corresponding eigenvectors and the
However, the ellipsoid of gyration is the more fundamental square root of their ratio determines the ratio of the major and minor
representation; it is readily obtained even when the Jacobian axes of the ellipsoid, which are colinear with the eigenvectors.
of the linkage is noninvertible. Also, while the matrix Y(8) Assuming the upper extremity may be modelled as a two-member
linkage with equal link lengths, without biarticuiar muscles, a
may never have zero eigenvalues, (assuming real links with necessary condition to achieve an end-point stiffness with equal
nonzero mass) the matrix W(d) may, because of the eigenvalues (hence a circular ellipsoid) is only satisfied at the point p
kinematics of the linkage. If the inertial behavior of the tip is on the workspace boundary as shown in (b). With biarticuiar muscles
written in the conventional (impedance) form: acting at equal moment arms about each joint an end-point stiffness
with equal eigenvalues and a circular ellipsoid may be achieved
p = M(0)V (21) throughout the region R shown in (c).

there exist locations in the workspace for which the eigen- example, when one tenses the muscles of the arm without
values of the tensor M(6) become infinite. Thus the end-point moving; the impedance of the limb increases.)
inertia tensor can not be defined for some linkage con- There are also considerably more skeletal muscles than
figurations. On the other hand the worst the eigenvalues of joints, even beyond the antagonist pairing required to permit
W(6) will do is go to zero, which is easier to deal with com- unidirectional muscle force to produce bidirectional joint
putationally. Again, a reminder of the fact that the difference torques. For example, the torque flexing the elbow joint (one
between impedance and admittance is fundamental. of the simpler joints in the primate upper extremity) is
generated by brachialis, brachioradialis, biceps capitus brevis,
Impedance Modulation Using Actuator Redundancies and biceps capitus longus. Does this complexity serve any
purpose? If the control of end-point impedance of the limb
It is also possible to modulate the position- and velocity- without feedback is considered it will seen that these apparent
dependent components of end-point impedance without actuator redundancies may have a functional role to play [13].
feedback by exploiting the intrinsic properties of the ac- Consider the simplified two-link model of the primate
tuators, and again apparent redundancies are useful. upper limb (forearm and hand treated as a single rigid body,
Although a muscle is by no means thermodynamically pin-jointed to the upper arm) moving in a horizontal plane as
conservative, it exhibits a static relation between force and shown in Fig. 3. In the absence of feedback, the static
length (for any given fixed level of neural input) similar to component of the total end-point impedance will solely be due
that of a mechanical spring, i.e., one which permits the to the spring-like properties of the individual muscles. For
definition of a potential function analogous to elastic energy.3 each muscle, a potential function may be defined, and the
Muscle force also exhibits a dependence on velocity similar to combined effect of multiple muscles is to define a total
a mechanical damper. It has been shown that the mechanical potential function (which could be determined by adding the
impedance of a single muscle may be modulated by neural potential functions of the individual muscles). The total
commands both in the presence and in the absence of neural potential at any point is invariant under coordinate trans-
feedback [7, 11, 12, 25, 26]. Simultaneously activating two or formations and the total potential function may be expressed
more muscles which oppose each other across a joint is one in any coordinate system by direct substitution.
strategy which permits impedance to be modulated in- Now, for simplicity, assume that the relations between
dependent of joint torque [15, 20]. (This is what happens, for muscle force and length and muscle length and joint rotation
Curiously, the force/length behaviour of most muscles is such that the co- result in a linear torque/angle relation for each muscle. First
energy integral is not defined and thus no compliance form is definable [29]: consider the monoarticular (single-joint) muscles which
Muscles are impedances, not admittances. generate torques about only a single joint: their combined

12/Vol. 107, MARCH 1985 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 effect is to define a diagonal stiffness tensor in relative controller has to solve, may in fact represent a solution to a
angular coordinates: problem: they may play a functional role in controlling the
interaction between the limb and the environment during
"7V Ks 0 ~P\~
dynamic events sufficiently rapid to limit the effectiveness of
= (22) feedback control.
.T2_ 0 Ke _P2_
Summary
Each of the terms Ks and Ke may vary. For example, the
stiffness about the human elbow can vary from about 1 In this part of the paper, techniques for implementing a
Nm/rad. to more than 200 Nm/rad [20, 36]. desired impedance on a manipulator were considered.
When this stiffness tensor is expressed in end-point Feedback control algorithms for imposing Cartesian im-
coordinates, because of the distortion due to the nonor- pedances up to second order on a general nonlinear
thogonality of the kinematic transformations the end-point manipulator were presented. Because care was taken to ask
stiffness will no longer be diagonal, but the range of end-point for a manipulator behaviour which is compatible with the
stiffnesses which could be achieved without feedback using fundamental mechanics of manipulation, (as outlined in Part
monoarticular muscles to change Ks and Ke is quite restricted. I) the need to solve the "inverse kinematics p r o b l e m " -
This is readily seen in the shape of the potential function generally regarded as fundamental to all robot control - was
corresponding to this stiffness. For small displacements the circumvented.
potential function is a quadratic form and its isopotential Techniques for modulating the end-point impedance of a
contours are ellipsoids which graphically represent the manipulator without recourse to feedback were also
directional character of the stiffness tensor (see Fig. 3(a)). discussed. Multiple actuators and "excess" linkage degrees of
To illustrate the nature of the problem, suppose it were freedom may also be used to modulate end-point impedance
desired to have the end-point equally stiff in all directions. and it is suggested that the apparent redundancies in the
This would correspond to a potential function with circular primate musculoskeletal system may in fact play an essential
isopotentials. However, given only single joint muscles, functional role in controlling interactive behavior. The
throughout the useful workspace a potential function with hypothesis that impedance modulation is one of the
circular isopotentials can not be achieved. For example, prominent strategies of natural movement control provides
assuming links of equal length and joint ranges of 0 to 90 the motivation for a research project to develop a cyber-
degrees for the shoulder and 0 to 180 degrees for the elbow, a netically controlled prosthesis which will give an amputee the
necessary condition to achieve circular isopotentials is only ability to change its impedance at will [2].
satisfied at one point (point p in Fig. 3(b)) on the boundary of The modulation of end-point impedance without feedback
the workspace (see Appendix III). This is because to specify a may also be important for industrial robots. Feedback loop
symmetric second-rank tensor such as stiffness in two transmission delays are not just a biological problem; It is
dimensions requires three parameters and the monoarticular widely recognized that computation time is one of the limiting
muscles provide only two. factors in the design of robot controllers. It could be argued
However, the biomechanical system abounds with that as computation becomes cheaper and faster, this problem
polyarticular muscles — muscles which generate torques about will disappear, but one reasonable way of describing
more than one joint. The biceps and triceps muscles of the manipulation is as a series of "collisions" with objects in the
upper arm cross both the elbow joint and the shoulder joint environment [10]. During a collision dynamic events take
and provide a mechanical coupling between shoulder and place extremely rapidly and any feedback controller may
elbow rotations which radically increases the range of stiff- encounter difficulties. Control of dynamic interaction
nesses which may be achieved without feedback. without feedback is an interesting alternative and is currently
For simplicity assume the same linear relation between under investigation [19].
muscle-generated torque and angle for both joints. Now, A feature of impedance control is that different controller
including the two-joint muscles, the stiffness tensor in relative actions (aimed at satisfying different task requirements) may
joint angle coordinates will have off-diagonal terms: be superimposed. For example, suppose that a desired end-
point position- and velocity-dependent behaviour is im-
Ks + Kt Kt plemented on a manipulator using a feedback control strategy
(23) as outlined above in equations (4) and (7). At the same time
Kt Ke + Kt kinematic redundancies in the manipulator are used to
modulate the end-point mobility. At any given end-point
The term Kt represents the contribution of the two-joint position, X, (which is determinable from the configuration, 6)
muscles and, like Ke and Ks, it may vary. Now suppose again the manipulator configuation may be chosen to best ap-
that it is desired to have the end-point equally stiff in all proximate a desired inertial behaviour (for example, the
directions. As a result of the two-joint muscles, as shown in mobility normal to a kinematic constraint surface may be
Appendix III, a potential field with circular isopotentials maximised). This configuration may then be used in the
could be achieved without feedback (by varying Ke, Ks, and feedback law which implements the position- and velocity-
Kt) throughout a much larger region in the workspace (region dependent behaviour. As the equations never require in-
R in Fig. 3(c)). In effect, the two-joint muscles provide a third version of the Jacobian, they can be applied to a manipulator
parameter with which to modulate the stiffness tensor. Note with kinematic redundancies. Note that this approach to end-
that this is not peculiar to the specific set of simplifying point control in the presence of kinematic redundancies is
assumptions made above: In general, the availability of significantly different from the use of a generalised
polyarticular muscles dramatically increases the range of end- pseudoinverse [35].
point impedances which could be achieved without feedback. Part III of this paper will discuss the application of im-
The point of this discussion is to demonstrate that im- pedance control.
pedance control is possible without depending on feedback
strategies, by using to advantage the intrinsic behavior of the
manipulator "hardware." Apparent redundancies in the Acknowledgments
musculoskeletal system, which are frequently seen as Portions of the work reported in this paper were supported
presenting a coordination problem which the biological by:

Journal of Dynamic Systems, Measurement, and Control MARCH 1985, Vol. 107/13

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 NSF Gant No. PFR 7917348 23 Makino, H., Furuya, N., Soma, K., and Chin, E., "Research and
Development of the SCARA Robot," Proceedings of the 4th International Con-
NIHR Grant No. GOO 820 0048, Department of Ed- ference on Production Engineering, Tokyo, 1980.
ucation 24 Morasso, P., "Spatial Control of Arm Movements," Experimental Brain
The Whitaker Health Sciences Fund Research, Vol. 42, 1981, pp. 223-227.
Polaroid Corporation 25 Nichols, T. R., "Soleus Muscle Stiffness and Its Reflex Control," Ph.D.
thesis, Harvard University, Cambridge, MA, 1974.
The John and Fannie Hertz Foundation
26 Nichols, T. R., and Houk, J. C , "Improvement in Linearity and Regula-
The Ralph E, Cross Fund tion of Stiffness that Results from Actions of Stretch Reflex," Journal of
American Can Company Neurophysiology, Vol. 39, 1976, pp. 119-142.
The TRW Foundation Faculty Fellowship 27 Paul, R. P., "Manipulator Cartesian Path Control," IEEE Transactions
This support is gratefully acknowledged. Portions of the work on Systems, Man, and Cybernetics, Vol. SMC-9, 1979, pp. 702-711.
28 Paul, R. P., Robot Manipulators: Mathematics, Programming, and Con-
were performed in the Eric P. and Evelyn E. Newman trol, MIT Press, Cambridge, Mass., 1981.
Laboratory for Biomechanics and Human Rehabilitation, the 29 Ruch, T. C , and Patton, H. E., Physiology and Biophysics, W. B.
Acoustics, Vibrations and Machine Dynamics Laboratory, Saunders Co., New York, 1966.
and the Laboratory for Manufacturing and Productivity. 30 Steifel, E. L., and Scheifele, G., Linear and Regular Celestial Mechanics,
Springer-Verlag, New York, 1971.
31 Takahashi, Y., Rabins, M. J., and Auslander, D. M., Control and
References Dynamic Systems, Addison-Wesley, Mass., 1970.
32 Taylor, R. H., "Planning and Execution of Straight Line Manipulator
1 Abend, W., Bizzi, E., and Morasso, P., "Human Arm Trajectory Forma- Trajectories," IBM Journal of Research and Development, Vol. 23, 1979, pp.
tion," Brain, Vol. 105, 1982, pp. 331-348. 424-436.
2 Abul-Haj, C. J., "The Design of an Upper Arm Prosthesis Simulator with 33 Whitaker, E. T., A Treatise on the Analytical Dynamics of Particles and
Variable Mechanical Impedance," S.M. thesis, Department of Mechanical Rigid Bodies, Cambridge University Press, 1904 and Dover, New York, 1944.
Engineering, MIT, Sept. 1981. 34 Whitney, D. E., "Resolved Motion Rate Control of Manipulators and
3 Andrews, J. R., and Hogan, N., "Impedance Control as a Framework for Human Prostheses," IEEE Transactions on Man-Machine Systems, Vol.
Implementing Obstacle Avoidance in a Manipulator," pp. 243-251 in: Control MMS-10, No. 2, June 1969, pp. 47-53.
of Manufacturing Processes and Robotic Systems, Eds. Hardt, D. E. and Book, 35 Whitney, D. E., "The Mathematics of Coordinated Control of Prosthetic
W. J., American Society of Mechanical Engineers, New York, 1983. Arms and Manipulators," ASME JOURNAL OF DYNAMIC SYSTEMS, MEASURE-
4 Asada, H., "A Geometrical Representation of Manipulator Dynamics MENT, AND CONTROL, Dec. 1972, pp. 303-309.
and its Application to Arm Design," pp. 1-8 in: Robotics Research and Ad- 36 Zahaiak, G. I., and Heyman, S. J., " A Quantitative Evaluation of the
vanced Applications, Ed. Book, W. J., American Society of Mechanical Frequency Response Characteristics of Active Human Skeletal Muscle in Vivo,"
Engineers, New York, 1982. ASME Journal of Biomechanical Engineering, Vol. 101, 1979, pp. 28-37.
5 Brady, M., Hollerbach, J., Johnson, T., Lozano-Perez, T., and Mason,
M., (Eds.), Robot Motion: Planning and Control, MIT Press, Cambridge,
1983. APPENDIX I
6 Cotter, S. L., "Nonlinear Feedback Control of Manipulator Endpoint
Impedance," S. M. thesis, Department of Mechanical Engineering, MIT, Sept., A Nonlinear Feedback Law for Impedance Control
1982.
7 Crago, P. E., Houk, J, C , and Hasan, Z., "Regulatory Actions of Assume that the desired end-point behavior to be imposed
Human Stretch Reflex," Journal of Neurophysiology, Vol. 39, 1976, pp. on the manipulator is given by:
925-935.
8 Crandall, S. H., Karnopp, D. C , Kurtz, E. F. Jr., and Pridmore-Brown, MdV/dt-B[Y0 - V] -ATX 0 - X ] =Fint
D. C , Dynamics of Mechanical and Electromechanical Systems, McGraw-Hill,
New York, 1968. Assume that an adequate model of the manipulator dynamics
9 Dubowsky, S., and Des Forges, D. T., "The Application of Model- is:
Referenced Adaptive Control to Robotic Manipulators," ASME JOURNAL OF
DYNAMIC SYSTEMS, MEASUREMENT AND CONTROL, Vol. 101, 1979, pp. 193-200. I(6)du/dt + C(0,to) + K(co) + S(0) = Tact + Tint
10 Goertz, R. C , "Manipulators Used for Handling Radioactive Materials," In this equation, 1(6) is the configuration-dependent inertia
Chapter 27 in: Human Factors in Technology, Ed. Bennett, E. M., McGraw-
Hill, 1963. tensor for the manipulator, C(0,to) are the inertial coupling
11 Gordon, A. M., Huxley, A. F., and Julian, F. J., "The Variation in terms (due to centrifugal and coriolis accelerations), K(u)
Isometric Tension with Sarcomere Length in Vertebrate Muscle Fibers," Jour- includes any velocity-dependent forces (e.g., frictional) and
nal of Physiology, Vol. 184, 1966, pp. 170-192. S(8) includes any static configuration-dependent forces (e.g.,
12 Hoffer, J. A., and Andreassen, S., "Limitations in the Servo-Regulation
of Soleus Muscle Stiffness in Premammillary Cats," in: Muscle Receptors and gravitational). Any actuator dynamics have been neglected.
Movement, Eds. Taylor, A. and Prochazka, A., Macmillan, London, 1981. The actuator forces (or torques) Tact are assumed to be the
13 Hogan, N., "Mechanical Impedance Control in Assistive Devices and control input to the manipulator.
Manipulators," Proceedings of the Joint Automatic Controls Conference, Vol. The equation for the desired behavior may be regarded as a
1, San Francisco, Aug. 1980.
14 Hogan, N., "Programmable Impedance Control of Industrial specification of the desired end-point acceleration which is to
Manipulators," Proceedings of the Conference on CAD/CAM in Mechanical result from an external force impressed on the manipulator
Engineering, MIT, Mar. 1982. admittance.
15 Hogan, N., "Adaptive Control of Mechanical Impedance by Coactivation
of Antagonist Muscles," IEEE Transactions on Automatic Control, Vol. orV/c// = M - , ^ [ X 0 - X ] + M ~ 1 B [ V 0 - V ] + M - 1 F i n t
AC-29, No. 8, August 1984, pp. 681-690.
16 Hogan, N., and Cotter, S. L., "Cartesian Impedance Control of a The corresponding acceleration in actuator coordinates is
Nonlinear Manipulator," pp. 121-128 in: Robotics Research and Advanced Ap- obtained by differentiating the kinematic transformations.
plications, Ed. Book, W. J., American Society of Mechanical Engineers, New
York, 1982. dX/dt = 3(d)du/dt + G(6,o>)
17 Hollerbach, J. M., " A Recursive Formulation of Lagrangian Manipulator where
Dynamics," IEEE Transactions on Systems, Man and Cybernetics, Vol. 10,
1980, pp. 730-736. G(6,u>) = [d{3(0)<a}/d6]oi
18 Horn, B. K. P., and Raibert, M. H., "Configuration Space Control,"
The Industrial Robot, June 1978, pp. 69-73. du/dt = J ~' (6) [dV/dt - G(6, co)]
19 Kleidon, M. A., "Modeling and Performance of a Pneumatic/Hydraulic
Actuator with Tunable Mechanical Impedance," S. M. thesis, Department of Each of the impedance terms in the desired end-point
Mechanical Engineering, MIT, Sept. 1983. behavior may be expressed in actuator coordinates using the
20 Lanman, J. M., "Movement and the Mechanical Properties of the Intact kinematic transformations
Human Elbow Joint," Ph.D. thesis, Department of Psychology, MIT, June
1980. K[Xo-X]=K[Xo-L(0)]
21 Luh, J. Y. S., Walker, M. W„ and Paul, R. P. C , "On-line Computa-
tional Scheme for Mechanical Manipulators," ASME JOURNAL OF DYNAMIC fi[Vo-V]=fl[Vo-J<0)«]
SYSTEMS, MEASUREMENT AND CONTROL, Vol. 102, pp. 69-76, June 1980.
22 Luh, J. Y. S., Walker, M. W., and Paul, R. P. C , "Resolved Accelera-
For the purposes of controller design, each of these terms may
tion Control of Mechanical Manipulators," IEEE Transactions on Automatic be regarded as a component of a desired feedback law relating
Control, Vol. AC-25, 1980, pp. 468-474. the control input Tact to the variables 6, co and Fint, which are

14/Vol. 107, MARCH 1985 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 assumed to be accessible measurements. The complete control dEk = dh' co = dp' V = dpJ(6)u
law is obtained by substitution. At any given configuration
Tact = 1(6)3-[(0)M-'K[XO-L(0)]+S(0) (position terms) h = J'(0)p
+ 1(0)3 -' (6)M-' B[V0 - J(0)co] + K(co) (velocity terms) These relations may be used to express the mobility in end-
+ 1(6)3 - ' (6)M-' Fint - J'(0)Fint (force terms) point coordinates.
- 1(0)3 - ' (0)G(8,w) + C(0,co) (inertial coupling terms) v=jco=jyj'p
Note that although this equation does require the inverse Denoting the end-point mobility by W(0)
Jacobian, it does not require inversion of the kinematic W(0) = 3Y3<
equations. Only the forward kinematic equations need be
computed. This will be important for those manipulators for V = W(6)p
which no explicit algebraic (closed form) solution to the in- The physical meaning of the mobility tensor is that if the
verse kinematic equations exists. system is at rest an applied force will produce an acceleration
equal to the force vector premultiplied by the mobility tensor.
At rest, d6/dt = 0 and hence:
APPENDIX II
d\/dt = 3u/dt
Generalized Inertial Systems and the Mobility Tensor
du/dt= Ydh/dt
Any mechanical linkage is a generalized inertial system. The
defining property of an inertial system is its ability to store From the generalized Hamiltonian [30]:
kinetic energy, defined as the integral of (generalized) velocity dh/dt = T- V0H
with respect to (generalized) momentum [8]. At any con-
figuration defined by the generalized coordinates the kinetic At rest, h = 0 henceH(h,0) = £ £ = 0 and V , H = 0 . Thus:
energy is a quadratic form in (generalized) momentum. dh/dt = T
Ek = 'Ah' Y(6)h d\/dt = 3Y3'F=W¥
From Hamilton's equations [30], the (generalized) velocity is As the mobility tensor is symmetric it may be diagonalized
the momentum gradient of the kinetic energy. by rotating the coordinate axes to coincide with its eigen-
vectors. A force applied in the direction of an eigenvector
H(h,6) = Ek(h,6) (when the system is at rest) results in an acceleration in the
d0/dt = u=VhH=Y(0)h same direction equal to the applied force multiplied by the
Kinetic energy is commonly confused with kinetic coenergy. corresponding eigenvalue. The eigenvalues represent the
The two are not identical and are related by a Legendre inverse of the apparent mass or inertia seen by the applied
transform [8]. force or torque.
Because the kinetic energy is a quadratic form in
Ek* = co' h - Ek = co' y ' co - Vi co1 Y-' YY-' co momentum, it may be represented graphically by an ellipsoid
Ek* = Vi co' Y-i(0)o>= !/2co'/(0)<o (see Fig. 2), the ellipsoid of gyration [33]. This may be
thought of as the set of all momenta which produce the same
At any configuration kinetic coenergy is a quadratic form in kinetic energy (an isokinetic contour in momentum space).
(generalized) velocity and its velocity gradient is the The lengths of the principal axes of the ellipsoid of gyration
(generalized) momentum [8]. are inversely proportional to the square roots of the eigen-
h = I(6)oi values, proportional to the square roots of the associated
apparent mass or inertia. The long direction of the ellipsoid of
For a generalized inertial system, Y is a symmetric, twice-
Fig. 2 is the direction of the greatest apparent inertia.
contravariant tensor. To distinguish it from its inverse, the
inertia tensor /, (symmetric, twice-covariant) Y will be termed In the general case when the system is not at rest the relation
the mobility tensor. between applied force and resulting motion is (in general)
A knowledge of the geometric relation between coordinate nonlinear and must be written in terms of a complete set of
frames is sufficient to transform any tensor from one frame state equations for the inertial system. A convenient set of
to another. As the joint angles are a set of generalized state variables are the Hamiltonian states, generalized
coordinates, for any configuration of the linkage of Fig. 2 the position (e.g., 0) and generalized momentum (h). The state
end-point coordinates are related to the joint angles via the and output equations are in the form of generalized ad-
kinematic transformations. mittance (see Part I) as follows.
State equations:
X = L(0)
dh/dt=- V 9 [!/2h'y(0)h]+J'(0)F
Differentiating these transformations yields the relation
between velocities (at any given configuration). dd/dt= Vh[Vih'Y(0)\i\= Y(0)h
dX/dt = V = J(0)co Output equations (position and velocity):
3(6) in these equations is the configuration-dependent X=L(0)
Jacobian. As the coordinate transformation does not store, V=3(0)Y(6)b
dissipate or generate energy, incremental changes in energy
are the same in all coordinate frames. This yields the relation
between forces in each coordinate frame. APPENDIX III
dEp = Td6 = F'dX = F' 3(6)d6 Effect of Actuator Redundancy on Range of Feasible Stiff-
ness
At any given configuration
The differential stiffness tensor in relative joint angle
T = 3'(0)F coordinates (Pi,p 2 ! due to the combined stiffnesses of
The same approach yields the relation between the monoarticular actuators, Ks, Ke and biarticular actuators Kt,
momenta in each coordinate frame. is:

Journal of Dynamic Systems, Measurement, and Control MARCH 1985, Vol. 107/15

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Ks + Kt Kt To achieve an isotropic end-point stiffness (for which the
corresponding potential function will have circular
Kt Ke+Kt Pi isopotentials) its eigenvalues must be equal. For simplicity
The transformation from relative joint angle coordinates assume each eigenvalue is unity.
(Pi,p 2 ) t 0 absolute joint angle coordinates (0i,0 2 l is: Kx=\
1 0' The corresponding stiffness tensor in absolute angle coor-
dinates is:
-1 1 Ko= J'AjtJ = J'J
Hence the stiffness tensor in absolute joint angle coordinates
is: L, L2 cos (d2 -&\)
Ko-
1 Ks+Kt Kt 1 0" L, L2 cos (82 — 0,) V
1 Kt Ke+Kt -1 1
To achieve an isotropic end-point stiffness it is necessary for
" Ks + Ke -Ke the actual joint coordinate stiffness to equal the desired joint
coordinate stiffness. Assuming L, =L2 = 1 it can be seen that
-Ke Kt + Ke in the absence of biarticular actuators, i.e., Kt = 0, this
The differential transformation from absolute joint angle condition is not satisfied except at:
coordinates [8],d2] to Cartesian end-point coordinates
02-0, =180°

dX{ -Li sin 0! -L2 sin 62 ~d6x point p in figure 3b. In contrast, given bi-articular actuators,
i.e., Kt^Q, isotropic stiffness can be achieved throughout the
dX2 Li cos 61 L2 cos d2 __ dd2 region R in Fig. 3(c) defined by:

dX = Jdd 9 O ° < 0 , - 0 , <180°

16/Vol. 107, MARCH 1985 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/ on 09/08/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use