195

Force Contro
Multimedia Contents 9. Force Control

Luigi Villani, Joris De Schutter

9.1 Background .......................................... 195

A fundamental requirement for the success of
9.1.1 From Motion Control
a manipulation task is the capability to handle
to Interaction Control ................... 196
the physical contact between a robot and the en- 9.1.2 From Indirect Force Control
vironment. Pure motion control turns out to be to Hybrid Force/Motion Control ...... 197
inadequate because the unavoidable modeling
errors and uncertainties may cause a rise of the 9.2 Indirect Force Control ............................ 198
contact force, ultimately leading to an unstable 9.2.1 Stiffness Control ........................... 198
behavior during the interaction, especially in the 9.2.2 Impedance Control....................... 201
presence of rigid environments. Force feedback
and force control becomes mandatory to achieve 9.3 Interaction Tasks ................................... 205
9.3.1 Rigid Environment ....................... 205
a robust and versatile behavior of a robotic sys-
9.3.2 Compliant Environment ................ 208
tem in poorly structured environments as well as
9.3.3 Task Specification......................... 209
safe and dependable operation in the presence of
9.3.4 Sensor-Based
humans. This chapter starts from the analysis of
Contact Model Estimation ............. 211
indirect force control strategies, conceived to keep
the contact forces limited by ensuring a suitable 9.4 Hybrid Force/Motion Control .................. 211
compliant behavior to the end effector, without 9.4.1 Acceleration-Resolved Approach ... 211
requiring an accurate model of the environment. 9.4.2 Passivity-Based Approach ............. 214
Then the problem of interaction tasks modeling 9.4.3 Velocity-Resolved Approach .......... 215
is analyzed, considering both the case of a rigid
environment and the case of a compliant environ- 9.5 Conclusions
ment. For the specification of an interaction task, and Further Reading ............................. 216
natural constraints set by the task geometry and 9.5.1 Indirect Force Control ................... 216
9.5.2 Task Specification......................... 217
artificial constraints set by the control strategy are
9.5.3 Hybrid Force/Motion Control.......... 217
established, with respect to suitable task frames.
This formulation is the essential premise to the Video-References......................................... 217
synthesis of hybrid force/motion control schemes.
References................................................... 218 Part A | 9.1

9.1 Background
Research on robot force control has flourished in the Since the early work on telemanipulation, the use
past three decades. Such a wide interest is motivated of force feedback was conceived to assist the human
by the general desire of providing robotic systems operator in the remote handling of objects with a slave
with enhanced sensory capabilities. Robots using force, manipulator. More recently, cooperative robot systems
touch, distance, and visual feedback are expected to au- have been developed where two or more manipula-
tonomously operate in unstructured environments other tors (viz. the fingers of a dexterous robot hand) are
than the typical industrial shop floor. to be controlled so as to limit the exchanged forces
 196 Part A Robotics Foundations

and avoid squeezing of a commonly held object. Force end-effector from the desired trajectory. On the other
control plays a fundamental role also in the achieve- hand, the control system reacts to reduce such devia-
ment of robust and versatile behavior of robotic systems tions. This ultimately leads to a build-up of the contact
in open-ended environments, providing intelligent re- force until saturation of the joint actuators is reached or
sponse in unforeseen situations and enhancing human– breakage of the parts in contact occurs.
robot interaction. The higher the environment stiffness and position
control accuracy are, the more easily a situation like
9.1.1 From Motion Control the one just described can occur. This drawback can be
to Interaction Control overcome if a compliant behavior is ensured during the
interaction. This compliant behavior can be achieved ei-
Control of the physical interaction between a robot ther in a passive or in an active fashion.
manipulator and the environment is crucial for the suc-
cessful execution of a number of practical tasks where Passive Interaction Control
the robot end-effector has to manipulate an object or In passive interaction control the trajectory of the robot
perform some operation on a surface. Typical exam- end-effector is modified by the interaction forces due to
ples in industrial settings include polishing, deburring, the inherent compliance of the robot. The compliance
machining or assembly. A complete classification of may be due to the structural compliance of the links,
possible robot tasks, considering also nonindustrial ap- joints, and end-effector, or to the compliance of the po-
plications, is practically infeasible in view of the large sition servo. Soft robot arms with elastic joints or links
variety of cases that may occur, nor would such a clas- are purposely designed for intrinsically safe interaction
sification be really useful to find a general strategy to with humans. In industrial applications, a mechanical
control the interaction with the environment. device with passive compliance, known as the remote
During contact, the environment may set constraints center of compliance (RCC) device [9.1], is widely
on the geometric paths that can be followed by the end- adopted. An RCC is a compliant end-effector mounted
effector, denoted as kinematic constraints. This situa- on a rigid robot, designed and optimized for peg-into-
tion, corresponding to the contact with a stiff surface, hole assembly operations.
is generally referred to as constrained motion. In other The passive approach to interaction control is
cases, the contact task is characterized by a dynamic very simple and cheap, because it does not require
interaction between the robot and the environment that force/torque sensors; also, the preprogrammed trajec-
can be inertial (as in pushing a block), dissipative (as tory of the end-effector must not be changed at ex-
in sliding on a surface with friction) or elastic (as in ecution time; moreover, the response of a passive
pushing against an elastically compliant wall). In all compliance mechanism is much faster than active repo-
these cases, the use of a pure motion control strategy for sitioning by a computer control algorithm. However,
controlling interaction is prone to failure, as explained the use of passive compliance in industrial applications
below. lacks flexibility, since for every robotic task a special-
Successful execution of an interaction task with the purpose compliant end-effector has to be designed and
environment by using motion control could be obtained mounted. Also, it can only deal with small position and
only if the task were accurately planned. This would in orientation deviations of the programmed trajectory. Fi-
turn require an accurate model of both the robot manip- nally, since no forces are measured, it can not guarantee
ulator (kinematics and dynamics) and the environment that high contact forces will never occur.
(geometry and mechanical features). A manipulator
model may be known with sufficient precision, but Active Interaction Control
a detailed description of the environment is difficult to In active interaction control, the compliance of the
obtain. robotic system is mainly ensured by a purposely de-
To understand the importance of task planning accu- signed control system. This approach usually requires
Part A | 9.1

racy, it is sufficient to observe that in order to perform the measurement of the contact force and moment,
a mechanical part mating with a positional approach the which are fed back to the controller and used to mod-
relative positioning of the parts should be guaranteed ify or even generate online the desired trajectory of the
with an accuracy of an order of magnitude greater than robot end-effector.
part mechanical tolerance. Once the absolute position Active interaction control may overcome the afore-
of one part is exactly known, the manipulator should mentioned disadvantages of passive interaction control,
guide the motion of the other with the same accuracy. but it is usually slower, more expensive, and more so-
In practice, the planning errors may give rise to phisticated. To obtain a reasonable task execution speed
a contact force and moment, causing a deviation of the and disturbance rejection capability, active interaction
 Force Control 9.1 Background 197

control has to be used in combination with some degree of impedance and admittance control are stiffness con-
of passive compliance [9.2]: feedback, by definition, trol and compliance control [9.9], respectively, where
always comes after a motion and force error has oc- only the static relationship between the end-effector po-
curred, hence some passive compliance is needed in sition and orientation deviation from the desired motion
order to keep the reaction forces below an acceptable and the contact force and moment is considered. Notice
threshold. that, in the robot control literature, the terms impedance
control and admittance control are often used to refer to
Force Measurements the same control scheme; the same happens for stiffness
For a general force-controlled task, six force compo- and compliance control. Moreover, if only the relation-
nents are required to provide complete contact force ship between the contact force and moment and the end-
information: three translational force components and effector linear and angular velocity is of interest, the
three torques. Often, a force/torque sensor is mounted corresponding control scheme is referred to as damping
at the robot wrist [9.3], but other possibilities exist, for control [9.10].
example, force sensors can be placed on the fingertips Indirect force control schemes do not require, in
of robotic hands [9.4]; also, external forces and mo- principle, measurements of contact forces and mo-
ments can be estimated via shaft torque measurements ments; the resulting impedance or admittance is typi-
of joint torque sensors [9.5, 6]. However, the majority cally nonlinear and coupled. However, if a force/torque
of the applications of force control (including indus- sensor is available, then force measurements can be
trial applications) is concerned with wrist force/torque used in the control scheme to achieve a linear and de-
sensors. In this case, the weight and inertia of the tool coupled behavior.
mounted between the sensor and the environment (i. e., Differently from indirect force control, direct force
the robot end-effector) is assumed to be negligible or control requires an explicit model of the interaction
suitably compensated from the force/torque measure- task. In fact, the user has to specify the desired motion
ments. The force signals may be obtained using strain and the desired contact force and moment in a con-
measurements, which results in a stiff sensor, or de- sistent way with respect to the constraints imposed
formation measurements (e.g., optically), resulting in by the environment. A widely adopted strategy be-
a compliant sensor. The latter approach has an advan- longing to this category is hybrid force/motion control,
tage if additional passive compliance is desired. which aims at controlling the motion along the uncon-
strained task directions and force (and moment) along
9.1.2 From Indirect Force Control the constrained task directions. The starting point is the
to Hybrid Force/Motion Control observation that, for many robotic tasks, it is possible to
introduce an orthogonal reference frame, known as the
Active interaction control strategies can be grouped into compliance frame [9.11] (or task frame [9.12]) which
two categories: those performing indirect force control allows one to specify the task in terms of natural and
and those performing direct force control. The main dif- artificial constrains acting along and about the three
ference between the two categories is that the former orthogonal axes of this frame. Based on this decompo-
achieve force control via motion control, without ex- sition, hybrid force/motion control allows simultaneous
plicit closure of a force feedback loop; the latter instead control of both the contact force and the end-effector
offer the possibility of controlling the contact force and motion in two mutually independent subspaces. Simple
moment to a desired value, thanks to the closure of selection matrices acting on both the desired and feed-
a force feedback loop. back quantities serve this purpose for planar contact
To the first category belongs impedance control (or surfaces [9.13], whereas suitable projection matrices
admittance control) [9.7, 8], where the deviation of the must be used for general contact tasks, which can also
end-effector motion from the desired motion due to the be derived from the explicit constraint equations [9.14–
interaction with the environment is related to the con- 16]. Several implementation of hybrid motion control
Part A | 9.1

tact force through a mechanical impedance/admittance schemes are available, e.g., based on inverse dynam-
with adjustable parameters. A robot manipulator under ics control in the operational space [9.17], passivity-
impedance (or admittance) control is described by an based control [9.18], or outer force control loops closed
equivalent mass–spring–damper system with adjustable around inner motion loops, typically available in indus-
parameters. This relationship is an impedance if the trial robots [9.2].
robot control reacts to the motion deviation by gener- If an accurate model of the environment is not
ating forces, while it corresponds to an admittance if available, the force control action and the motion con-
the robot control reacts to interaction forces by impos- trol action can be superimposed, resulting in a parallel
ing a deviation from the desired motion. Special cases force/position control scheme. In this approach, the
 198 Part A Robotics Foundations

force controller is designed so as to dominate the erated along the constrained task directions in order to
motion controller; hence, a position error would be tol- ensure force regulation [9.19].

9.2 Indirect Force Control

To gain insight into the problems arising at the inter- 6  1 vector xe D . pTe 'Te /T , where 'e is a set of Eu-
action between the end-effector of a robot manipulator ler angles extracted from Re . Hence, a desired end-
and the environment, it is worth analyzing the effects of effector position and orientation can be assigned in
a motion control strategy in the presence of a contact terms of a vector xd , corresponding to the position of
force and moment. To this aim, assume that a ref- the origin pd and the rotation matrix Rd of a desired
erence frame ˙e is attached to the end-effector, and frame ˙d . The end-effector error can be denoted as
let pe denote the position vector of the origin and Re xde D xd  xe , and the corresponding velocity error,
the rotation matrix with respect to a fixed base frame. assuming a constant xd , can be expressed as Pxde D
The end-effector velocity is denoted by the 6  1 twist Pxe D A1 . 'e /v e , with
vector v e D .PpTe !Te /T where pP e is the translational ve-
locity and !e the angular velocity, and can be computed  
I 0
from the n  1 joint velocity vector qP using the linear A. 'e / D ;
0 T. 'e /
mapping
where I is the 3  3 identity matrix, 0 is a 3  3 null
v e D J. q/Pq : (9.1)
matrix, and T is the 3  3 matrix of the mapping
!e D T. 'e /'P e , depending on the particular choice of
The matrix J is the 6  n end-effector geometric Ja-
the Euler angles.
cobian. For simplicity, the case of nonredundant non-
Consider the motion control law
singular manipulators is considered; therefore, n D 6
and the Jacobian is a square nonsingular matrix. The
force f e and moment me applied by the end-effector hc D AT . 'e /KP xde  KD v e C . q/ ; (9.3)
to the environment are the components of the wrench
he D . f Te mTe /T . corresponding to a simple proportional–derivative
It is useful to consider the operational space formu- (PD) + gravity compensation control in the operational
lation of the dynamic model of a rigid robot manipula- space, where KP and KD are symmetric and positive-
tor in contact with the environment definite 6  6 matrices.
In the absence of interaction with the environment
ƒ. q/vP e C . q; qP /v e C . q/ D hc  he ; (9.2) (i. e., when he D 0), the equilibrium v e D 0, xde D 0
for the closed-loop system, corresponding to the de-
where sired position and orientation for the end-effector, is
asymptotically stable. The stability proof is based on
ƒ. q/ D .JH. q/1 JT /1 the positive-definite Lyapunov function

is the 6  6 operational space inertia matrix, 1 1

V D v Te ƒ. q/v e C xde KP xde ;
P 1 2 2
. q; qP / D JT C. q; qP /J1  ƒ. q/JJ
whose time derivative along the trajectories of the
is the wrench including centrifugal and Coriolis effects, closed-loop system is the negative semidefinite function
and . q/ D JT g. q/ is the wrench of the gravitational
effects; H. q/, C. q; qP / and g. q/ are the corresponding
Part A | 9.2

VP D v Te KD v e : (9.4)
quantities defined in the joint space. The vector hc D
JT  is the equivalent end-effector wrench correspond-
ing to the input joint torques . In the presence of a constant wrench he , using a similar
Lyapunov argument, a different asymptotically stable
9.2.1 Stiffness Control equilibrium can be found, with a nonnull xde . The new
equilibrium is the solution of the equation
In the classical operational space formulation, the end-
effector position and orientation is described by a AT . 'e /KP xde  he D 0 ;
 Force Control 9.2 Indirect Force Control 199

which can be written in the form in the case of an infinitesimal twist displacement •xde
defined as
xde D K1
P A . 'e /he ;
T
(9.5)      
•pde Ppde pP
or, equivalently, as •xde D D dt D  e dt ;
• de !de !e
he D AT . 'e /KP xde : (9.6)
where Ppde D pP d  pP e is the time derivative of the posi-
Equation (9.6) shows that in the steady state the end- tion error pde D pd  pe and !de D !d  !e is the
effector, under a proportional control action on the angular velocity error. Equation (9.7) shows that the
position and orientation error, behaves as a six-degree- actual stiffness matrix is AT . 'e /KP A1 . 'e /, which
of-freedom (DOF) spring in respect of the external depends on the end-effector orientation through the vec-
force and moment he . Thus, the matrix KP plays the tor 'e , so that, in practice, the selection of the stiffness
role of an active stiffness, meaning that it is possible parameters is quite difficult.
to act on the elements of KP so as to ensure a suit- This problem can be overcome by defining a ge-
able elastic behavior of the end-effector during the ometrically consistent active stiffness, with the same
interaction. Analogously, (9.5) represents a compliance structure and properties as ideal mechanical springs.
relationship, where the matrix K1 P plays the role of
an active compliance. This approach, consisting of as- Mechanical Springs
signing a desired position and orientation and a suitable Consider two elastically coupled rigid bodies A and B
static relationship between the deviation of the end-ef- and two reference frames ˙a and ˙b , attached to A
fector position and orientation from the desired motion and B, respectively. Assuming that at equilibrium
and the force exerted on the environment, is known as frames ˙a and ˙b coincide, the compliant behavior
stiffness control. near the equilibrium can be described by the linear map-
The selection of the stiffness/compliance parame- ping
ters is not easy, and strongly depends on the task to be  
Kt Kc
executed. A higher value of the active stiffness means hbb D K•xbab D •xbab ; (9.8)
a higher accuracy of the position control at the expense KTc Ko
of higher interaction forces. Hence, if it is expected to
meet some physical constraint in a particular direction, where hbb is the elastic wrench applied to body B, ex-
the end-effector stiffness in that direction should be pressed in frame B, in the presence of an infinitesimal
made low to ensure low interaction forces. Conversely, twist displacement •xbab of frame ˙a with respect to
along the directions where physical constraints are not frame ˙b , expressed in frame B. The elastic wrench
expected, the end-effector stiffness should be made high and the infinitesimal twist displacement in (9.8) can
so as to follow closely the desired position. This allows also be expressed equivalently in frame ˙a , since ˙a
discrepancies between the desired and achievable posi- and ˙b coincide at equilibrium. Therefore, hbb D hab
tions due to the constraints imposed by the environment and •xbab D •xaab ; moreover, for the elastic wrench ap-
to be resolved without excessive contact forces and mo- plied to body A, haa D Kt •xaba D hbb being •xaba D
ments. •xbab . This property of the mapping (9.8) is known as
It must be pointed out, however, that a selective port symmetry.
stiffness behavior along different directions cannot be In (9.8), K is the 6  6 symmetric positive-semidef-
effectively assigned in practice on the basis of (9.6). inite stiffness matrix. The 3  3 matrices Kt and Ko ,
This can easily be understood by using the classi- called respectively the translational stiffness and ro-
cal definition of a mechanical stiffness for two bodies tational stiffness, are also symmetric. It can be shown
connected by a 6-DOF spring, in terms of the linear that, if the 3  3 matrix Kc , called the coupling stiffness
mapping between the infinitesimal twist displacement is symmetric, there is maximum decoupling between
Part A | 9.2

of the two bodies at an unloaded equilibrium and the rotation and translation. In this case, the point corre-
elastic wrench. sponding to the coinciding origins of the frames ˙a
In the case of the active stiffness, the two bod- and ˙b is called the center of stiffness. Similar defini-
ies are, respectively, the end-effector, with the attached tions and results can be formulated for the case of the
frame ˙e , and a virtual body, attached to the desired compliance matrix C D K1 . In particular, it is possi-
frame ˙d . Hence, from (9.6), the following mapping ble to define a center of compliance in the case that the
can be derived off-diagonal blocks of the compliance matrix are sym-
metric. The center of stiffness and compliance do not
he D AT . 'e /KP A1 . 'e /•xde ; (9.7) necessarily coincide.
200 Part A Robotics Foundations

There are special cases in which no coupling ex- punov, a suitable potential elastic energy function must
ists between translation and rotation, i. e., a relative be defined.
translation of the bodies results in a wrench correspond- For simplicity, it is assumed that the coupling stiff-
ing to a pure force along an axis through the center ness matrix is zero. Hence, the potential elastic energy
of stiffness; also, a relative rotation of the bodies re- can be computed as the sum of a translational potential
sults in a wrench that is equivalent to a pure torque energy and a rotational potential energy.
about an axis through the centers of stiffness. In these The translational potential energy can be defined as
cases, the center of stiffness and compliance coincide.
Mechanical systems with completely decoupled behav- 1
Vt D pTde K0Pt pde ; (9.9)
ior are, e.g., the remote center of compliance (RCC) 2
devices.
Since Kt is symmetric, there exists a rotation ma- with
trix Rt with respect to the frame ˙a D ˙b at equilib-
1 1
rium, such that Kt D Rt  t RTt , and  t is a diagonal K0Pt D Rd KPt RTd C Re KPt RTe ;
matrix whose diagonal elements are the principal trans- 2 2
lational stiffnessess in the directions corresponding to where KPt is a 3  3 symmetric positive-definite matrix.
the columns of the rotation matrix Rt , known as the The use of K0Pt in lieu of KPt in (9.9) guarantees that
principal axes of translational stiffness. Analogously, the potential energy is port symmetric also in the case
Ko can be expressed as Ko D Ro  o RTo , where the of finite displacements. Matrices K0Pt and KPt coincide
diagonal elements of  o are the principal rotational at equilibrium (i. e., when Rd D Re ) and in the case of
stiffnesses about the axes corresponding to the columns isotropic translational stiffness (i. e., when KPt D kPt I).
of rotation matrix Ro , known as the principal axes of The computation of the power VP t yields
rotational stiffness. Moreover, assuming that the ori-
gins of ˙a and ˙b at equilibrium coincide with the VP t D PpeT
de f
t C !de m
t ;
e eT e
center of stiffness, the expression Kc D Rc  c RTc can
be found, where the diagonal elements of  c are the where Ppede is the time derivative of the posi-
principal coupling stiffnesses along the directions cor- tion displacement pede D RTe . pd  pe /, while !ede D
responding to the columns of the rotation matrix Rc , RTe . !d  !e /. The vectors f
t
e
and e
t are, respec-
known as the principal axes of coupling stiffness. In tively, the elastic force and moment applied to the end-
sum, a 6  6 stiffness matrix can be specified, with re- effector in the presence of the finite position displace-
spect to a frame with origin in the center of stiffness, in ment pede . These vectors have the following expres-
terms of the principal stiffness parameters and principal sions when computed in the base frame
axes.
Notice that the mechanical stiffness defined by (9.8) f
t D K0Pt pde m
t D K00Pt pde ; (9.10)
describes the behavior of an ideal 6-DOF spring which
stores potential energy. The potential energy function with
of an ideal stiffness depends only on the relative posi-
tion and orientation of the two attached bodies and is 1
port symmetric. A physical 6-DOF spring has a pre- K00Pt D S.pde /Rd KPt RTd ;
2
dominant behavior similar to the ideal one, but never-
theless it always has parasitic effects causing energy where S./ is the skew-symmetric operator performing
dissipation. the vector product. The vector h
t D . f T
t mT
t /T is
the elastic wrench applied to the end-effector in the
Geometrically Consistent Active Stiffness presence of a finite position displacement pde and
To achieve a geometrically consistent 6-DOF active a null orientation displacement. The moment m
t is
Part A | 9.2

stiffness, a suitable definition of the proportional con- null in the case of isotropic translational stiffness.
trol action in control law (9.3) is required. This control To define the rotational potential energy, a suitable
action can be interpreted as the elastic wrench applied definition of the orientation displacement between the
to the end-effector, in the presence of a finite displace- frames ˙d and ˙e has to be adopted. A possible choice
ment of the desired frame ˙d with respect to the end- is the vector part of the unit quaternion fde ; ede g that
effector frame ˙e . Hence, the properties of the ideal can be extracted from matrix Red D RTe Rd . Hence, the
mechanical stiffness for small displacements should be orientation potential energy has the form
extended to the case of finite displacements. Moreover,
to guarantee asymptotic stability in the sense of Lya- Vo D 2eT
de KPo de ;
e
(9.11)
 Force Control 9.2 Indirect Force Control 201

where KPo is a 3  3 symmetric positive-definite matrix. this case, it can be shown that the principal axes of
The function Vo is port symmetric because ede D ded . rotational stiffness cannot be set arbitrarily but must co-
The computation of the power VP o yields incide with those of the end-effector frame.
Stiffness control with a geometrically consistent ac-
VP o D !eT
de m
o ;
e
tive stiffness can be defined using the control law

where hc D h
 KD v e C . q/ ; (9.15)

m
o D K0Po de ; (9.12) with h
in (9.13). The asymptotic stability about the
equilibrium in the case he D 0 can be proven using the
with Lyapunov function

K0Po D 2ET .de ; de /Re KPo RTe ; 1

V D v Te ƒ. q/v e C Vt C Vo ;
2
and E.de ; de / D de I  S. de /. The above equations
show that a finite orientation displacement de D RTe ede with Vt and Vo given in (9.9) and (9.11), respectively,
produces an elastic wrench h
o D .0T mT
o /T which is whose time derivative along the trajectories of the
equivalent to a pure moment. closed-loop system, in case the frame ˙d is motion-
Hence, the total elastic wrench in the presence less, has the same expression as in (9.4). When he ¤
of a finite position and orientation displacement of 0, a different asymptotically stable equilibrium can
the desired frame ˙d with respect to the end-effector be found, corresponding to a nonnull displacement of
frame ˙e can be defined in the base frame as the desired frame ˙d with respect to the end-effector
frame ˙e . The new equilibrium is the solution of the
h
D h
t C h
o : (9.13) equation h
D he .
Stiffness control allows to keep the interaction force
where h
t and h
o are computed according to (9.10) and moment limited at the expense of the end-effector
and (9.12), respectively. position and orientation error, with a proper choice of
Using (9.13) for the computation of the elastic the stiffness matrix, without the need of a force/torque
wrench in the case of an infinitesimal twist displace- sensor. However, in the presence of disturbances (e.g.,
ment •xede near the equilibrium, and discarding the high- joint friction) which can be modeled as an equiva-
order infinitesimal terms, yields the linear mapping lent end-effector wrench, the adoption of low values
  for the active stiffness may produce large deviations
KPt 0 with respect to the desired end-effector position and
he D KP •xde D
e e
•xede : (9.14)
0 KPo orientation, also in the absence of interaction with the
environment.
Therefore, KP represents the stiffness matrix of an ideal
spring with respect to a frame ˙e (coinciding with ˙d 9.2.2 Impedance Control
at equilibrium) with the origin at the center of stiff-
ness. Moreover, it can be shown, using definition (9.13), Stiffness control is designed to achieve a desired static
that the physical/geometrical meaning of the principal behavior of the interaction. In fact, the dynamics of
stiffnesses and of the principal axes for the matri- the controlled system depends on that of the robot
ces KPt and KPo are preserved also in the case of large manipulator, which is nonlinear and coupled. A more
displacements. demanding objective may be that of achieving a de-
The above results imply that the active stiffness ma- sired dynamic behavior for the end-effector, e.g., that of
trix KP can be set in a geometrically consistent way a second-order mechanical system with six degrees of
with respect to the task at hand. freedom, characterized by a given mass, damping, and
Part A | 9.2

Notice that geometrical consistency can also be en- stiffness, known as mechanical impedance.
sured with different definitions of the orientation error The starting point to pursue this goal may be the ac-
in the potential orientation energy (9.11), for example, celeration-resolved approach used for motion control,
any error based on the angle/axis representation of Rde which is aimed at decoupling and linearizing the non-
can be adopted (the unit quaternion belongs to this cat- linear robot dynamics at the acceleration level via an
egory), or, more generally, homogeneous matrices or inverse dynamics control law. In the presence of inter-
exponential coordinates (for the case of both position action with the environment, the control law
and orientation errors). Also, the XYZ Euler angles ex-
tracted from the matrix Rde could be used; however, in hc D ƒ. q/˛ C . q; qP /Pq C . q/ C he (9.16)
 202 Part A Robotics Foundations

cast into the dynamic model (9.2) results in stability of the equilibrium in the case he D 0 can be
proven by considering the Lyapunov function
vP e D ˛ ; (9.17)
1
V D v eT
de KM v de C Vt C Vo ;
e
(9.22)
where ˛ is a properly designed control input with the 2
meaning of an acceleration referred to the base frame.
PN T v e , with
N Te vP ee C R
Considering the identity vP e D R
where Vt and Vo are defined in (9.9) and (9.11), respec-
e e tively, and whose time derivative along the trajectories
  of system (9.21) is the negative semidefinite function
N e D Re
R
0
;
0 Re VP D v eT
de KD v de :
e

the choice When he ¤ 0, a different asymptotically stable equilib-

PN T v e rium can be found, corresponding to a nonnull displace-
N T ˛e C R
˛DR e e e (9.18) ment of the desired frame ˙d with respect to the end-
effector frame ˙e . The new equilibrium is the solution
gives of the equation h
D he .
In case ˙d is constant, (9.21) has the meaning of
vP ee D ˛e ; (9.19) a true 6-DOF mechanical impedance if KM is chosen as
 
where the control input ˛e has the meaning of an ac- mI 0
celeration referred to the end-effector frame ˙e . Hence, KM D ;
0 M
setting
where m is a mass and M is a 33 inertia tensor, and KD
˛e D vP ed C K1
M .KD v de C h
 he / ;
e e e
(9.20) is chosen as a block-diagonal matrix with 3  3 blocks.
The physically equivalent system is a body of mass m
the following expression can be found for the closed- with inertia tensor M with respect to a frame ˙e at-
loop system tached to the body, subject to an external wrench hee .
This body is connected to a virtual body attached to
KM vP ede C KD v ede C he
D hee ; (9.21) frame ˙d through a 6-DOF ideal spring with stiffness
matrix KP and is subject to viscous forces and moments
where KM and KD are 6  6 symmetric and positive- with damping KD . The function V in (9.22) represents
definite matrices, vP ede D vP ed  vP ee , v ede D v ed  v ee , vP ed the total energy of the body: the sum of the kinetic and
and v ed are, respectively, the acceleration and the ve- potential elastic energy.
locity of a desired frame ˙d and he
is the elastic A block diagram of the resulting impedance control
wrench (9.13); all the quantities are referred to the end- is sketched in Fig. 9.1. The impedance control com-
effector frame ˙e . putes the acceleration input as in (9.18) and (9.20) on
The above equation describing the dynamic behav- the basis of the position and orientation feedback as
ior of the controlled end-effector can be interpreted as well as the force and moment measurements. Then, the
a generalized mechanical impedance. The asymptotic inverse dynamics control law computes the torques for
the joint actuators  D JT hc with hc in (9.16). This con-
trol scheme, in the absence of interaction, guarantees
pd , Rd he
υd Manipulator that the end-effector frame ˙e asymptotically follows
Impedance α Inverse τ q
υ· d control dynamics
and
q· the desired frame ˙d . In the presence of contact with the
environment
environment, a compliant dynamic behavior is imposed
Part A | 9.2

on the end-effector, according to the impedance (9.21),

and the contact wrench is bounded at the expense
of a finite position and orientation displacement be-
tween ˙d and ˙e . Differently from stiffness control,
pe, Re
υe Direct a force/torque sensor is required for the measurement
kinematics of the contact force and moment.
In the case that the force/torque sensor is not avail-
able, the measure of the external wrench he cannot
Fig. 9.1 Impedance control be used in the controller and thus the configuration-
 Force Control 9.2 Indirect Force Control 203

independent impedance behaviour (9.21) cannot be presence of interaction, while the contact wrench is null
obtained anymore. However, a desired impedance be- when the end-effector moves in free space.
haviour can be still achieved with the control law The disturbances acting on the robot manipulator
and the unmodeled dynamics (joint friction, modeling
hc D ƒ. q/vP d C. q; qP /v d CKD v de Ch
C. q/ errors, etc.) may be taken into account by introducing
an additive term on the right-hand side of the dynamic
(9.23)
model of the robot manipulator (9.2), corresponding to
an equivalent disturbance wrench acting on the end-
in place of (9.16), where KD is a 6  6 positive-definite effector. This term produces an additive acceleration
matrix and h
is the elastic wrench (9.13). The result- disturbance  e on the right-hand side of (9.19). There-
ing closed-loop equation is fore, using the control law (9.20), the following closed-
loop impedance equation can be found
ƒ. q/vP de C .. q; qP / C KD /v de C h
D he ;
KM vP ede C KD v ede C he
D hee C KM  e : (9.25)
representing an impedance behaviour which preserves
the actual operational space inertia matrix ƒ. q/ of the The tuning procedure for the impedance parameters
robot. In the above equation, the centrifugal and Cori- can be set up starting from the linearized model that
olis wrench  . q; qP /v de is required to preserve the can be computed from (9.25) in the case of infinitesimal
mechanical properties of a configuration-dependent in- displacements, i. e.,
ertia and to prove the stability similarly to (9.21). In
case ˙d is constant, control law (9.23) reduces to the KM •Rxede C KD •Pxede C .KP C K/•xede
stiffness control law (9.15). D K•xedo C KM  e ; (9.26)
Pioneering experiments on stiffness control and
impedance control and without without force sensing where (9.24) and the equality •xeeo D •xede C •xedo have
are presented in VIDEO 684 . VIDEO 686 reports been used. The above equation is valid both for con-
experiments on impedance control based on a geomet- strained (K ¤ 0) and for free motion (K D 0).
rically consistent active stiffness. It is evident that suitable dynamics of the position
and orientation errors can be set by suitably choosing
Implementation Issues the matrix gains KM , KD , and KP . This task is easier
The selection of good impedance parameters ensuring under the hypothesis that all the matrices are diagonal,
a satisfactory behavior is not an easy task. In fact, the resulting in a decoupled behavior for the six compo-
dynamics of the closed-loop system is different in free nents of the infinitesimal twist displacement. In this
space and during interaction. The control objectives are case, the transient behavior of each component can be
different as well, since motion tracking and disturbance set, e.g., by assigning the natural frequency and damp-
rejection must be ensured in free space, while, during ing ratio with the relations
the interaction, the main goal is achieving a suitable s
compliant dynamic behavior for the end-effector. No- < kP C k 1 kD
tice also that the dynamics of the controlled system !n D ; D p :
kM 2 kM .kP C k/
during the interaction depends on the dynamics of the
environment. Hence, if the gains are chosen so that a given natural
To gain insight into these problems, assume that frequency and damping ratio are ensured during the in-
the interaction of the end-effector with the environment teraction (i. e., for k ¤ 0), a smaller natural frequency
can be approximated by that derived from an ideal 6- with a higher damping ratio will be obtained when the
DOF spring connecting end-effector frame ˙e to the end-effector moves in free space (i. e., for k D 0). As for
environment frame ˙o . Therefore, according to (9.8), the steady-state performance, the end-effector error for
Part A | 9.2

the elastic wrench exerted by the end-effector on the the generic component is
environment, in the presence of an infinitesimal twist
displacement of ˙e with respect to ˙o , can be com- k kM
puted as •xde D •xdo C 
.kP C k/ kP C k

hee D K•xeeo ; (9.24) and the corresponding interaction force is

where ˙e and ˙o coincide at equilibrium and K is kP k kM k

hD •xdo  :
a stiffness matrix. The above model holds only in the kP C k kP C k
204 Part A Robotics Foundations

pd , Rd pc, Rc he Fig. 9.2 Impedance control

υd Impedance υc Pos and
α Inverse τ Manipulator q with inner motion control
orient and
υ· d control υ· c control
dynamics
environment q· loop (admittance control)

pe, Re
υe Direct
kinematics

The above relations show that, during interaction, the A block diagram of the resulting scheme is sketched
contact force can be made small at the expense of in Fig. 9.2. It is evident that, in the absence of in-
a large position error in steady state, as long as the ac- teraction, the compliant frame ˙c coincides with the
tive stiffness kP is set low with respect to the stiffness desired frame ˙d and the dynamics of the position and
of the environment k, and vice versa. However, both the orientation error, as well as the disturbance rejection ca-
contact force and the position error also depend on the pabilities, depend only on the gains of the inner motion
external disturbance ; in particular, the lower kP , the control loop. On the other hand, the dynamic behav-
higher the influence of  on both •xde and h. Moreover, ior in the presence of interaction is imposed by the
a low active stiffness kP may result in a large position er- impedance gains (9.27).
ror also in the absence of interaction (i. e., when k D 0). The control scheme of Fig. 9.2 is also known as ad-
mittance control because, in (9.27), the measured force
Admittance Control (the input) is used to compute the motion of the compli-
A solution to this drawback can be devised by separat- ant frame (the output), given the motion of the desired
ing motion control from impedance control as follows. frame; a mapping with a force as input and a position
The motion control action is purposefully made stiff so or velocity as output corresponds to a mechanical ad-
as to enhance disturbance rejection but, rather than en- mittance. Vice versa, (9.21), mapping the end-effector
suring tracking of the desired end-effector position and displacement (the input) from the desired motion tra-
orientation, it ensures tracking of a reference position jectory into the contact wrench (the output), has the
and orientation resulting from the impedance control meaning of a mechanical impedance. Experiments on
action. In other words, the desired position and orien- admittance control are reported in VIDEO 685 .
tation, together with the measured contact wrench, are
input to the impedance equation which, via a suitable Simplified Schemes
integration, generates the position and orientation to be The inverse dynamics control is model based and
used as a reference for the motion control. requires modification of current industrial robot con-
To implement this solution, it is worth introducing trollers, which are usually equipped with indepen-
a reference frame other than the desired frame ˙d . This dent proportional–integral (PI) joint velocity controllers
frame is referred to as the compliant frame ˙c , and is with very high bandwidth. These controllers are able
specified by the quantities pc , Rc , v c , and vP c that are to decouple the robot dynamics to a large extent, espe-
computed from pd , Rd , v d , and vP d and the measured cially in the case of slow motion, and to mitigate the
wrench hc , by integrating the equation effects of external forces on the manipulator motion if
the environment is sufficiently compliant. Hence, the
KM vP cdc C KD v cdc C hc
D hc ; (9.27)
closed-loop dynamics of the controlled robot can be ap-
Part A | 9.2

where hc
is the elastic wrench in the presence of a fi- proximated by
nite displacement between the desired frame ˙d and the
compliant frame ˙c . Then, a motion control strategy, qP D qP r
based on inverse dynamics, is designed so that the end- in joint space, or equivalently
effector frame ˙e is taken to coincide with the compli-
ant frame ˙c . To guarantee the stability of the overall vP e D v r (9.28)
system, the bandwidth of the motion controller should
be higher than the bandwidth of the impedance con- in the operational space, where qP r and v r are the control
troller. signals for the inner velocity motion loop generated by
 Force Control 9.3 Interaction Tasks 205

a suitably designed outer control loop. These control corresponding to a compliant behavior of the end-effec-
signals are related by tor characterized by a damping KD and a stiffness KP .
In the case KP D 0, the resulting scheme is known as
qP r D J1 . q/v r : damping control.
Alternatively, an admittance-type control scheme
The velocity v r , corresponding to a velocity-resolved can be adopted, where the motion of a compliant
control, can be computed as frame ˙c can be computed as the solution of the dif-
ferential equation
 e 
v er D v ed C K1
D h
 hee ;
KD v cdc C hc
D hce
where the control input has been referred to the end- in terms of the position pc , orientation Rc , and velocity
effector frame, KD is a 6  6 positive-definite matrix twist v c , where the inputs are the motion variables of
and h
is the elastic wrench (9.13) with stiffness ma- the desired frame ˙d and the contact wrench hce . The
trix KP . The resulting closed-loop equation is motion variables of ˙c are then input to an inner po-
sition and velocity controller. In the case KD D 0, the
KD v ede C he
D hee ; resulting scheme is known as compliance control.

9.3 Interaction Tasks

Indirect force control does not require explicit knowl- of the wrist force/torque sensor or of the tool (e.g.,
edge of the environment, although to achieve a satisfac- an end-effector mounted on an RCC device).
tory dynamic behavior the control parameters have to be  Local deformation of the contact surfaces may oc-
tuned for a particular task. On the other hand, a model cur during the interaction, producing distributed
of the interaction task is required for the synthesis of contact areas (e.g., the case of a soft contact surface
direct force control algorithms. of the tool or of the environment).
An interaction task is characterized by complex  Static and dynamic friction may occur in the case of
contact situations between the manipulator and the en- non ideally smooth contact surfaces.
vironment. To guarantee proper task execution, it is
necessary to have an analytical description of the in- The design of the interaction control and the perfor-
teraction force and moment, which is very demanding mance analysis are usually carried out under simplify-
from a modeling viewpoint. ing assumptions. The following two cases are consid-
A real contact situation is a naturally distributed ered:
phenomenon in which the local characteristics of the
contact surfaces as well as the global dynamics of 1. The robot and the environment are perfectly rigid
the manipulator and environment are involved. In and purely kinematics constraints are imposed by
detail: the environment,
2. The robot is perfectly rigid, all the compliance in
 The environment imposes kinematic constraints on the system is localized in the environment, and the
the end-effector motion, due to one or more con- contact wrench is approximated by a linear elastic
tacts of different type, and a reaction wrench arises model.
when the end-effector tends to violate the con-
straints (e.g., the case of a robot sliding a rigid tool In both cases, frictionless contact is assumed. It is ob-
on a frictionless rigid surface). vious that these situations are only ideal. However, the

Part A | 9.3

The end-effector, while being subject to kinematic robustness of the control should be able to cope with
constraints, may also exert a dynamic wrench on situations where some of the ieal assumptions are re-
the environment, in the presence of environment dy- laxed. In that case the control laws may be adapted to
namics (e.g., the case of a robot turning a crank, deal with nonideal characteristics.
when the crank dynamics is relevant, or a robot
pushing against a compliant surface). 9.3.1 Rigid Environment
 The contact wrench may depend on the structural
compliance of the robot, due to the finite stiffness The kinematic constraints imposed by the environ-
of the joints and links of the manipulator, as well as ment can be represented by a set of equations that the
206 Part A Robotics Foundations

variables describing the end-effector position and ori- where Sf denotes a weighted pseudoinverse of the ma-
entation must satisfy; since these variables depend on trix Sf , i. e.,
the joint variables through the direct kinematic equa-  1 T
tions, the constraint equations can also be expressed in Sf D STf WSf Sf W ; (9.34)
the joint space as
where W is a suitable weighting matrix.
. q/ D 0 : (9.29) Notice that, while the range space of the matrix Sf
in (9.32) is uniquely defined by the geometry of the
The vector  is an m1 function, with m < n, where n is
contact, the matrix Sf itself is not unique; also, the con-
the number of joints of the manipulator, assumed to be
straint equations (9.29), the corresponding Jacobian J
nonredundant; without loss of generality, the case n D 6
is considered. Constraints of the form (9.29), involv- as well as the pseudoinverse Sf and the vector are not
ing only the generalized coordinates of the system, are uniquely defined.
known as holonomic constraints. The case of time-vary- In general, the physical units of measure of the ele-
ing constraints of the form . q; t/ D 0 is not considered ments of are not homogeneous and the columns of the
here but can be analyzed in a similar way. Moreover, matrix Sf , as well as of the matrix Sf , do not necessar-
only bilateral constraints expressed by equalities of the ily represent homogeneous entities. This may produce
form (9.29) are of concern; this means that the end- invariance problems in the transformation (9.33) if he
effector always keeps contact with the environment. represents a measured wrench that is subject to distur-
The analysis presented here is known as kinetostatic bances and, as a result, may have components outside
analysis. the range space of Sf . If a physical unit or a reference
It is assumed that the vector (9.29) is twice differ- frame is changed, the matrix Sf undergoes a transforma-
entiable and that its m components are linearly indepen- tion; however, the result of (9.33) with the transformed
dent at least locally in a neighborhood of the operating pseudoinverse in general depends on the adopted phys-
point. Hence, differentiation of (9.29) yields ical units or on the reference frame. The reason is that
the pseudoinverse is the weighted least-squares solution
J . q/Pq D 0 ; (9.30) of a minimization problem based on the norm of the
vector he  Sf . q/ , and the invariance can be guaran-
where J . q/ D @=@q is the m  6 Jacobian of . q/, teed only if a physically consistent norm of this vector
known as the constraint Jacobian. By virtue of the is used. In the ideal case that he is in the range space
above assumption, J . q/ is of rank m at least locally of Sf , there is a unique solution for in (9.33), regard-
in a neighborhood of the operating point. less of the weighting matrix, and hence the invariance
In the absence of friction, the generalized inter- problem does not appear.
action forces are represented by a reaction wrench A possible solution consists of choosing Sf so that
that tends to violate the constraints. This end-effector its columns represent linearly independent wrenches.
wrench produces reaction torques at the joints that can This implies that (9.31) gives he as a linear combi-
be computed using the principle of virtual work as nation of wrenches and is a dimensionless vector.
A physically consistent norm on the wrench space can
 e D JT . q/ ; be defined based on the quadratic form hTe K1 he , which
has the meaning of an elastic energy if K is a positive-
where is an m  1 vector of Lagrange multipliers. The
definite matrix corresponding to a stiffness. Hence, the
end-effector wrench corresponding to  e can be com-
choice W D K1 can be made for the weighting matrix
puted as
of the pseudoinverse.
he D JT . q/ e D Sf . q/ ; (9.31) Notice that, for a given Sf , the constraint Jaco-
bian can be computed from (9.32) as J . q/ D STf J. q/;
where moreover, the constraint equations can be derived by in-
Part A | 9.3

tegrating (9.30).
Sf D JT . q/JT . q/ : (9.32) Using (9.1) and (9.32), the equality (9.30) can be
rewritten in the form
From (9.31) it follows that he belongs to the m-
dimensional vector space spanned by the columns of J . q/J1 . q/J. q/Pq D STf v e D 0 ; (9.35)
the 6  m matrix Sf . The inverse of the linear transfor-
mation (9.31) is computed as which, by virtue of (9.31), is equivalent to

D Sf . q/he ; (9.33) hTe v e D 0 : (9.36)

 Force Control 9.3 Interaction Tasks 207

Equation (9.36) represents the kinetostatic relation- the configuration of the robot in contact with the en-
ship, known as reciprocity, between the ideal reaction vironment can be described in terms of a .6  m/  1
wrench he (belonging to the so-called force-controlled vector r of independent variables. From the implicit
subspace) and the end-effector twist that obeys the con- function theorem, this vector can be defined as
straints (belonging to the so-called velocity-controlled
subspace). The concept of reciprocity, expressing the rD . q/ ; (9.40)
physical fact that, in the hypothesis of rigid and fric-
tionless contact, the wrench does not cause any work where . q/ is any .6  m/  1 twice-differentiable vec-
against the twist, is often confused with the concept of tor function such that the m components of . q/ and
orthogonality, which makes no sense in this case be- the n  m components of . q/ are linearly independent
cause twists and wrenches belong to different spaces. at least locally in a neighborhood of the operating point.
Equations ((9.35)) and (9.36) imply that the veloc- This means that the mapping (9.40), together with the
ity-controlled subspace is the reciprocal complement of constraint (9.29), is locally invertible, with inverse de-
the m-dimensional force-controlled subspace, identified fined as
by the range of matrix Sf . Hence, the dimension of the
velocity-controlled subspace is 6  m and a 6  .6  m/
q D . r/ ; (9.41)
matrix Sv can be defined, whose columns span the ve-
locity-controlled subspace, i. e.,
where . r/ is a 6  1 twice-differentiable vector func-
v e D Sv . q/ ; (9.37) tion. Equation (9.41) explicitly provides all the joint
vectors which satisfy the constraint (9.29). Moreover,
where  is a suitable .6  m/  1 vector. From (9.35) the joint velocity vectors that satisfy (9.30) can be com-
and (9.37) the following equality holds puted as

STf . q/Sv . q/ D 0 I (9.38) qP D J . r/Pr ;

moreover, the inverse of the linear transformation (9.37) where J . r/ D @=@r is a 6.6m/ full-rank Jacobian
can be computed as matrix. Therefore, the following equality holds

 D Sv . q/v e ; (9.39) J . q/J . r/ D 0 ;

where Sv denotes a suitable weighted pseudoinverse of which can be interpreted as a reciprocity condition be-
the matrix Sv , computed as in (9.34). tween the subspace of the reaction torques spanned by
Notice that, as for the case of Sf , although the range the columns of the matrix JT and the subspace of the
space of the matrix Sv is uniquely defined, the choice constrained joint velocities spanned by the columns of
of the matrix Sv itself is not unique. Moreover, the the matrix J .
columns of Sv are not necessarily twists and the scalar  Rewriting the above equation as
may have nonhomogeneous physical dimensions. How-
ever, in order to avoid invariance problems analogous
J . q/J. q/1 J. q/J . r/ D 0 ;
to that considered for the case of Sf , it is convenient
to select the columns of Sv as twists so that the vec-
tor  is dimensionless; moreover, the weighting matrix and taking into account (9.32) and (9.38), the matrix Sv
used to compute the pseudoinverse in (9.39) can be set can be expressed as
as W D M, being M a 6  6 inertia matrix; this corre-
sponds to defining a norm in the space of twists based Sv D J. q/J . r/ ; (9.42)
Part A | 9.3

on the kinetic energy. It is worth observing that the

transformation matrices of twists and wrenches, corre- which, by virtue of (9.40) and (9.41), it can be equiva-
sponding to a change of reference frame, are different; lently expressed as a function of either q or r.
however, if twists are defined with angular velocity The matrices Sf and Sv , and their pseudoinverse Sf
on top and translational velocity on bottom, then their and Sv are known as selection matrices. They play
transformation matrix is the same as for wrenches. a fundamental role for the task specification, i. e., the
The matrix Sv may also have an interpretation in specification of the desired end-effector motion and in-
terms of Jacobians, as for Sf in (9.32). Due to the pres- teraction forces and moments, as well as for the control
ence of m independent holonomic constraints (9.29), synthesis.
208 Part A Robotics Foundations

9.3.2 Compliant Environment (and infinitesimal twist displacements) that are in the
range space of Sv . The twists Pv v are denoted as twists
In many applications, the interaction wrench between of freedom while the twists .I  Pv /v are denoted as
the end-effector and a compliant environment can twists of constraint.
be approximated by an ideal elastic model of the In the hypothesis of frictionless contact, the interac-
form (9.24). However, since the stiffness matrix K tion wrench between the end-effector and the environ-
is positive definite, this model describes a fully con- ment is restricted to a force-controlled subspace defined
strained case, when the environment deformation co- by the m-dimensional range space of a matrix Sf , as for
incides with the infinitesimal twist displacement of the the rigid environment case, i. e.,
end-effector. In general, however, the end-effector mo-
tion is only partially constrained by the environment he D Sf D hs ; (9.47)
and this situation can be modeled by introducing a suit-
able positive-semidefinite stiffness matrix. where is an m  1 dimensionless vector. Premulti-
The stiffness matrix describing the partially con- plying both sides of (9.44) by STf and using (9.43),
strained interaction between the end-effector and the (9.45), (9.46), and (9.47) yields
environment can be computed by modeling the en-
vironment as a couple of rigid bodies, S and O,
STf •xeo D STf C Sf ;
connected through an ideal 6-DOF spring of compli-
ance C D K1 . Body S is attached to a frame ˙s and
is in contact with the end-effector; body O is attached where the identity STf Pv D 0 has been exploited. There-
to a frame ˙o which, at equilibrium, coincides with fore, the following elastic model can be found
frame ˙s . The environment deformation about the equi-
librium, in the presence of a wrench hs , is represented he D Sf D K0 •xeo ; (9.48)
by the infinitesimal twist displacement •xso between
frames ˙s and ˙o , which can be computed as where K0 D Sf .STf CSf /1 STf is the positive-semidefi-
nite stiffness matrix corresponding to the partially con-
•xso D C hs : (9.43) strained interaction.
If the compliance matrix C is adopted as a weight-
All the quantities hereafter are referred to frame ˙s but
the superscript s is omitted for brevity. ing matrix for the computation of Sf , then K0 can be
For the considered contact situation, the end- expressed as
effector twist does not completely belong to the ideal
velocity subspace, corresponding to a rigid environ- K0 D Pf K ; (9.49)
ment, because the environment can deform. Therefore,
the infinitesimal twist displacement of the end-effector where Pf D Sf Sf . Being Pf Pf D Pf , matrix Pf is a pro-
frame ˙e with respect to ˙o can be decomposed as jection matrix that filters out all the end-effector
wrenches that are not in the range space of Sf .
•xeo D •xv C •xf ; (9.44) The compliance matrix for the partially constrained
interaction cannot be computed as the inverse of K0 ,
where •xv is the end-effector infinitesimal twist dis- since this matrix is of rank m < 6. However, us-
placement in the velocity-controlled subspace, defined ing (9.46), (9.43), and (9.47), the following equality can
as the 6  m reciprocal complement of the force-con- be found
trolled subspace, while •xf is the end-effector infinites-
imal twist displacement corresponding to the environ-
•xf D C0 he ;
ment deformation. Hence,

•xv D Pv •xeo ; where the matrix

Part A | 9.3

(9.45)
•xf D .I  Pv /•xeo D .I  Pv /•xso ; (9.46)
C0 D .I  Pv /C ; (9.50)
where Pv D Sv Svand Sv and Sv are defined as in the
rigid environment case. Being Pv Pv D Pv , the matrix Pv of rank 6  m, is positive semidefinite. If the stiffness
is a projection matrix that filters out all the end-effector matrix K is adopted as a weighting matrix for the com-
twists (and infinitesimal twist displacements) that are putation of Sv , then the matrix C0 has the noticeable
not in the range space of Sv . Moreover, I  Pv is a pro- expression C0 D C  Sv .STv K Sv /1 STv , showing that C0
jection matrix that filters out all the end-effector twists is symmetric.
 Force Control 9.3 Interaction Tasks 209

9.3.3 Task Specification and the desired velocities:

An interaction task can be assigned in terms of a desired

 A nonzero linear velocity along the zt -axis
end-effector wrench hd and twist v d . In order to be con-
 An arbitrary angular velocity about the zt -axis.
sistent with the constraints, these vectors must lie in the The task continues until a large reaction force in the
force- and velocity-controlled subspaces, respectively. zt direction is measured, indicating that the peg has hit
This can be guaranteed by specifying vectors d and  d the bottom of the hole, not represented in the figure.
and computing hd and v d as Hence, the matrices Sf and Sv can be chosen as

hd D Sf d ; v d D Sv  d ; 0 1 0 1
1 0 0 0 0 0
where Sf and Sv have to be suitably defined on the basis B0 1 0 0C B0 0C
B C B C
of the geometry of the task, and so that invariance with B0 0 0 0C B1 0C
Sf B
D B0
C; Sv D B C;
respect to the choice of the reference frame and change B 0 1 0C
C
B0
B 0CC
of physical units is guaranteed. @0 0 0 1A @0 0A
Many robotic tasks have a set of orthogonal ref- 0 0 0 0 0 1
erence frames in which the task specification is very
easy and intuitive. Such frames are called task frames
where the columns of Sf have the dimensions of
or compliance frames. An interaction task can be spec-
wrenches and those of Sv have the dimensions of twists,
ified by assigning a desired force/torque or a desired
defined in the task frame, and they transform accord-
linear/angular velocity along/about each of the frame
ingly when changing the reference frame. The task
axes. The desired quantities are known as artificial con-
frame can be chosen either attached to the end-effector
straints because they are imposed by the controller;
or to the environment.
these constraints, in the case of rigid contact, are
complementary to those imposed by the environment,
Turning a Crank
known as natural constraints.
The goal of this task is turning a crank with an idle han-
Some examples of task frame definition and task
dle. The handle has two degrees of motion freedom,
specification are given below.
corresponding to the rotation about the zt -axis and to
the rotation about the rotation axis of the crank. Hence
Peg-in-Hole
the dimension of the velocity-controlled subspace is
The goal of this task is to push the peg into the hole
6  m D 2, while the dimension of the force-controlled
while avoiding wedging and jamming. The peg has two
subspace is m D 4. The task frame can be assumed as
degrees of motion freedom, hence the dimension of the
in the Fig. 9.4, attached to the crank. The task can be
velocity-controlled subspace is 6  m D 2, while the di-
achieved by assigning the following desired forces and
mension of the force-controlled subspace is m D 4. The
torques:
task frame can be chosen as shown in Fig. 9.3 and the
task can be achieved by assigning the following desired  Zero forces along the xt and zt axes
forces and torques:  Zero torques about the xt and yt axes.
 Zero forces along the xt and yt axes and the following desired velocities:
 Zero torques about the xt and yt axes,

zt

yt
xt
Part A | 9.3

zt
yt
xt

Fig. 9.3 Insertion of a cylindrical peg into a hole Fig. 9.4 Turning a crank with an idle handle
 210 Part A Robotics Foundations

 A nonzero linear velocity along the yt -axis Hence, the matrices Sf and Sv can be chosen as
 An arbitrary angular velocity about the zt -axis. 0 1 0 1
0 0 0 1 0 0
Hence, the matrices Sf and Sv can be chosen as B0 0 0C B0 1 0C
B C B C
B1 0 0C B0 0 0C
0 1 0 1 B
Sf D B C B C:
1 0 0 0 0 0 C; Sv D B C
B0 B0 1 0C B0 0 0C
B 0 0 0C
C
B1
B 0CC @0 0 1A @0 0 0A
B0 1 0 0C B0 0C
Sf D B
B0
C; Sv D B C; 0 0 0 0 0 1
B 0 1 0C
C
B0
B 0CC
@0 0 0 1A @0 0A The elements of the 6  6 stiffness matrix K0 , corre-
0 0 0 0 0 1 sponding to the partially constrained interaction of the
end-effector with the environment, are all zero except
referred to the task frame. In this case, the task those of the 3  3 principal minor K0m formed by rows
frame is fixed with respect to the crank, but in mo- 3; 4; 5 and columns 3; 4; 5 of K0 , which can be com-
tion with respect both the end-effector frame (fixed puted as
to the handle) and to the base frame of the robot. 0 11
Hence, the matrices Sf and Sv are time variant when c3;3 c3;4 c3;5
referred either to the end-effector frame or to the base K0m D @c4;3 c4;4 c4;5 A ;
frame. c5;3 c5;4 c5;5

Sliding a Block on a Planar Elastic Surface where ci;j D cj;i are the elements of the compliance ma-
The goal of this task is to slide a prismatic block over trix C.
a planar surface along the xt -axis, while pushing with
General Contact Model
a given force against the elastic planar surface. The ob-
The task frame concept has proven to be very useful
ject has three degrees of motion freedom, hence the
for the specification of a variety of practical robotic
dimension of the velocity-controlled subspace is 6 
tasks. However, it only applies to task geometries with
m D 3 while the dimension of the force-controlled
limited complexity, for which separate control modes
subspace is m D 3. The task frame can be chosen at-
can be assigned independently to three pure transla-
tached to the environment, as shown in Fig. 9.5, and
tional and three pure rotational directions along the axes
the task can be achieved by assigning the desired
of a single frame. For more complex situations, as in
velocities:
the case of multiple-point contact, a task frame may
 A nonzero velocity along the xt -axis
not exist and more complex contact models have to be
 A zero velocity along the yt -axis
adopted. A possible solution is represented by the vir-
 A zero angular velocity about the zt -axis,
tual contact manipulator model, where each individual
contact is modeled by a virtual kinematic chain between
the manipulated object and the environment, giving the
and the desired forces and torques:
manipulated object (instantaneously) the same motion
 A nonzero force along the zt -axis
freedom as the contact. The velocities and force kine-
 Zero torques about the xt and yt axes.
matics of the parallel manipulator, formed by the virtual
manipulators of all individual contacts, can be derived
using the standard kinematics procedures of real manip-
ulators and allow the construction of bases for the twist
and wrench spaces of the total motion constraint.
A more general approach, known as constraint-
Part A | 9.3

based task specification opens up new applications in-

volving complex geometries and/or the use of multiple
zt sensors (force/torque, distance, visual sensors) for con-
yt trolling different directions in space simultaneously.
xt The concept of task frame is extended to that of mul-
tiple feature frames. Each of the feature frames makes
it possible to model part of the task geometry using
Fig. 9.5 Sliding of a prismatic object on a planar elastic translational and rotational directions along the axes of
surface a frame; also, part of the constraints is specified in each
 Force Control 9.4 Hybrid Force/Motion Control 211

these situations is exactly why force control is used. The

a) robustness of the force controller increases if the matri-
xt
xr ces Sf and Sv can be continuously updated, using mo-
υxt
yr υ tion and/or force measurements, during task execution.
θ
In detail, a nominal model is assumed to be avail-
yt
able; when the contact situation evolves differently
υyt from what the model predicts, the measured and pre-
dicted motion and force will begin to deviate. These
small incompatibilities can be measured and can then
be used to adapt the model online, using algorithms de-
rived from classical state-space prediction–correction
b)
estimators such as the Kalman filter.
xt xr Figure 9.6 reports an example of error between
yr fyt nominal and measured motion and force variables, typ-
fxt ical of a two-dimensional contour-following task. The
yt θ
fyt orientation of the contact normal changes if the envi-
ronment is notplanar. Hence an angular error  appears
between the nominal contact normal, aligned to the yt -
axis of the task frame (the frame with axes xt and yt ),
and the real contact normal, aligned to the yr -axis of the
Fig.9.6a,b Estimation of orientation error: (a) velocity- real task frame (the frame with axes xr and yr ). This
based approach, (b) force-based approach angle can be estimated with either velocity or force
measurements only:
of the feature frames. The total model and the total set
of constraints are achieved by collecting the partial task
 Velocity-based approach: the executed linear veloc-
ity v , which is tangent to the real contour (aligned
and constraints descriptions, each expressed in the indi-
to the xr -axis), does not completely lie along the
vidual feature frames.
xt -axis, but has a small component vyt along the
yt -axis. The orientation error  can then be approx-
9.3.4 Sensor-Based imated by  D tan1 .vyt =vxt /.
Contact Model Estimation
 Force-based approach: the measured (ideal) contact
force f does not completely lie along the nomi-
The task specification relies on the definition of the ve-
nal normal direction, aligned to the yt -axis, but has
locity-controlled subspace and of the force-controlled
a small component fxt along the xt -axis. The ori-
subspace assuming that an accurate model of the con-
entation error  can then be approximated by  D
tact is available all the time. On the other hand, in
tan1 .fxt =fyt /.
most practical implementations, the selection matri-
ces Sf and Sv are not exactly known, however many The velocity-based approach is disturbed by me-
interaction control strategies turn out to be rather robust chanical compliance in the system; the force-based
against modeling errors. In fact, to cope reliably with approach is disturbed by contact friction.

9.4 Hybrid Force/Motion Control

The aim of hybrid force/motion control is to split up ing the nonlinear robot dynamics at the acceleration
simultaneous control of both end-effector motion and level, via an inverse dynamics control law. In the pres-
Part A | 9.4

contact forces into two separate decoupled subprob- ence of interaction with the environment, a complete
lems. In the following, the main control approaches in decoupling between the force- and velocity-controlled
the hybrid framework are presented for the cases of both subspaces is sought. The basic idea is that of designing
rigid and compliant environments. a model-based inner control loop to compensate for the
nonlinear dynamics of the robot manipulator and de-
9.4.1 Acceleration-Resolved Approach couple the force and velocity subspaces; hence an outer
control loop is designed to ensure disturbance rejec-
As for the case of motion control, the acceleration- tion and tracking of the desired end-effector force and
resolved approach is aimed at decoupling and lineariz- motion.
 212 Part A Robotics Foundations

Rigid Environment force multipliers through (9.31). On the other hand, the
In the case of a rigid environment, the external wrench 6  m independent components of the end-effector ve-
can be written in the form he D Sf . The force multi- locity twist can be controlled through hc via (9.53).
plier can be computed from (9.2) by multiplying both An inverse-dynamics inner control loop can be de-
sides of (9.2) by the weighted pseudo-inverse Sf with signed by choosing the control wrench hc as
weight ƒ1 . q/ and using the time derivative of the last hc D ƒ. q/Sv ˛v C Sf f C . q; qP / C ƒ. q/SP v  ;
equality (9.35). This yields
(9.55)

D Sf . q/ Œhc  . q; qP / C ƒf . q/SP Tf v e ; (9.51) where ˛v and f are properly designed control inputs.
Substituting (9.55) into (9.53) and (9.51) yields
where ƒf . q/ D .STf ƒ1 Sf /1 and . q; qP / D  qP C . P D ˛ ;
Replacing again (9.51) into (9.2) gives
D f ;
ƒ. q/vP e C Sf ƒf . q/SP Tf v e D Nf . q/Œhc  . q; qP / ; showing that control law (9.55) allows complete de-
(9.52) coupling between the force- and velocity-controlled
subspaces.
with Nf D I  Sf Sf . Notice that Nf Nf D Nf and Nf Sf D It is worth noticing that, for the implementation
0, hence the 6  6 matrix Nf is a projection matrix that of the control law (9.55), the constraint (9.29) as well
filters out all the end-effector wrenches lying in the as (9.40) defining the vector of the configuration vari-
range of Sf . These correspond to wrenches that tend ables for the constrained system are not required, pro-
to violate the constraints. Equation (9.52) represents vided that the matrices Sf and Sv are known or estimated
a set of six second-order differential equations whose online. In these cases, the task can easily be assigned by
solution, if initialized on the constraints, automatically specifying a desired force, in terms of the vector d .t/,
satisfy (9.29) at all times. and a desired velocity, in terms of the vector  d .t/;
The reduced-order dynamics of the constrained sys- moreover, a force/velocity control is implemented.
tem is described by 6  m second-order equations that The desired force d .t/ can be achieved by setting
are obtained by premultiplying both sides of (9.52) by f D d .t/ ; (9.56)
the matrix STv and substituting the acceleration vP e with
but this choice is very sensitive to disturbance forces,
vP e D Sv P C SP v  : since it contains no force feedback. Alternative choices
are
The resulting equations are
h i f D d .t/ C KP Πd .t/  .t/ ; (9.57)
ƒv . q/P D STv hc  . q; qP /  ƒ. q/SP v  ; (9.53)
or
where Zt
f D d .t/ C KI Πd . /  . / d ; (9.58)
ƒv D STv ƒSv
0
and the identities (9.38) and where KP and KI are suitable positive-definite matrix
STv Nf D STv gains. The proportional feedback is able to reduce the
force error due to disturbance forces, while the integral
have been used. Moreover, expression (9.51) can be action is able to compensate for constant bias distur-
rewritten as bances.
h i The implementation of force feedback requires the
D Sf . q/ hc  . q; qP /  ƒ. q/SP v  ;
Part A | 9.4

(9.54) computation of the force multiplier  from the mea-

surement of the end-effector wrench he , which can be
where the identity SP Tf Sv D STf SP v has been exploited. achieved by using (9.33).
Equation (9.54) reveals that the force multiplier vec- Velocity control is achieved by setting
tor instantaneously depends also on the applied input
wrench hc . Hence, by suitably choosing hc , it is possi- ˛ D P d .t/ C KP Œ d .t/  .t/
ble to directly control the m independent components Zt
of the end-effector wrench that tend to violate the con- C KI Πd . /  . / d ; (9.59)
straints; these components can be computed from the m 0
 Force Control 9.4 Hybrid Force/Motion Control 213

where KP and KI are suitable matrix gains. It into (9.17) and premultiplying both sides of the result-
is straightforward to show that asymptotic tracking ing equation once by Sv and once by STf , the following
of  d .t/ and P d .t/ is ensured with exponential conver- decoupled equations can be derived
gence for any choice of positive-definite matrices KP
and KI . P D ˛ ; (9.63)
The computation of the vector  from the available R D f : (9.64)
measurements can be achieved using (9.39), where the
end-effector twist is computed from joint position and Hence, by choosing ˛ according to (9.59) as for the
velocity measurements through (9.1). rigid environment case, asymptotic tracking of a desired
Equations (9.57) or (9.58) and (9.59) represent the velocity  d .t/ and acceleration P d .t/ is ensured, with
outer control loop ensuring force/velocity control and exponential convergence. The control input f can be
disturbance rejection. chosen as
When (9.29) and (9.40) are known, the matrices Sf 
f D R d .t/ C KD P d .t/  .t/
P
and Sv can be computed according to (9.32) and (9.42)
and a force/position control can be designed by speci- C KP Πd .t/  .t/ ; (9.65)
fying a desired force d .t/ and a desired position rd .t/. ensuring asymptotic tracking of a desired force tra-
Force control can be designed as above, while posi- P d .t/;
jectory ( d .t/; R d .t/) with exponential conver-
tion control can be achieved by setting gence for any choice of positive-definite matrices KD
and KP .
˛ D rRd .t/ C KDr ŒPrd .t/  .t/ C KPr Œrd .t/  r.t/ : Differently from the rigid environment case, feed-
back of P is required for the implementation of the
Asymptotic tracking of rd .t/, rP d .t/ and rR d .t/ is en- force control law (9.65). This quantity could be com-
sured with exponential convergence for any choice of puted from end-effector wrench measurements he as
positive-definite matrices KDr and KPr . The vector r
required for position feedback can be computed from P D S hP e :
f
joint position measurements via (9.40).
However, since the wrench measurement signal is often
P is often replaced by
noisy, the feedback of
Compliant Environment
In the case of a compliant environment, according to the P D S K0 J. q/Pq ;
f (9.66)
decomposition (9.44) of the end-effector displacement,
the end-effector twist can be decomposed as where joint velocities are measured using tachometers
or computed from joint positions via numerical differ-
v e D Sv  C C0 Sf P ; (9.60) entiation and K0 is the positive-semidefinite stiffness
matrix (9.49) describing the partially constrained inter-
where the first term is a twist of freedom, the second action. For the computation of (9.66), only knowledge
term is a twist of constraint, the vector  is defined as (or an estimate) of K0 is required, and not that of the
in (9.42), and C0 is defined in (9.50). Assuming a con- full stiffness matrix K. Also, the implementation of the
stant contact geometry and compliance, i. e., SP v D 0, control law (9.62) requires knowledge (or an estimate)
CP 0 D 0, and SP f D 0, a similar decomposition holds in of the compliance matrix C0 of the partially constrained
terms of acceleration interaction and not that of the full compliance matrix C.
If the contact geometry is known, but only an es-
vP e D Sv P C C0 Sf R : (9.61) timate of the stiffness/compliance of the environment
is available, the control law (9.62), with (9.65), may
An inverse-dynamics control law (9.16) can be still guarantee the convergence of the force error, if
adopted, resulting in the closed loop (9.17), where ˛ a constant desired force d is assigned. In this case, the
Part A | 9.4

is a properly designed control input. control law (9.62) has the form
In view of the acceleration decomposition (9.61),
˛ D Sv ˛ C b
C0 Sf f ;
the choice
0
where bC D .I  Pv /bC and bC is an estimate of the com-
˛ D Sv ˛ C C0 Sf f  (9.62)
pliance matrix. Hence, (9.63) still holds, while, in lieu
of (9.64), the following equality can be found
allows decoupling of the force control from veloc-
ity control. In fact, substituting (9.61) and (9.62) R D Lf f ;

 214 Part A Robotics Foundations

where Lf D .STf C Sf /1 STf b

C Sf is a nonsingular matrix. Premultiplying both sides of (9.68) by the ma-
Thus, the force- and velocity-controlled subspaces re- trix Sv , the following expression for the reduced-order
main decoupled and the velocity control law (9.59) dynamics is achieved
does not need to be modified. On the other hand, if
ƒv . q/Ps C  v . q; qP /s C K s D 0 ; (9.69)
the feedback of the time derivative of is computed
PO with  v D STv  . q; qP /Sv C STv ƒ. q/SP v ; it can easily be
using (9.66), only an estimate can be obtained. Us-
ing (9.66), (9.60) and (9.48), the following equality can shown that the skew symmetry of the matrix ƒ. P q/ 
be found 2. q; qP / implies that the matrix ƒ P v . q/  2 v . q; qP / is
skew symmetric as well.
PO P On the other hand, premultiplying both sides
D L1
f :
of (9.68) by the matrix STf ƒ1 . q/, the following ex-
pression for the force dynamics can be found
Therefore, computing the force control law f as
PO h i
in (9.65) with a constant d , P and KD D
in lieu of f  D Sf . q/  0 . q; qP /  .Sv /T K s ;
KD I, the dynamics of the closed-loop system is
(9.70)
R C KD
P C Lf KP D Lf KP d ; being Sf the weighted pseudo-inverse of Sf with
weight ƒ1 . q/. The above equation shows that the
showing that exponential asymptotic stability of the force multiplier instantaneously depends on the con-
equilibrium D d can be ensured, also in the pres- trol input f but also on the error s in the velocity-
ence of the uncertain matrix Lf , with a suitable choice controlled subspace.
of the gains KD and KP . The asymptotic stability of the reduced-order sys-
tem (9.69) can be ensured with the choices
9.4.2 Passivity-Based Approach
P r D P d C ˛ ; (9.71)
The passivity-based approach exploits the passivity r D  d C ˛x ; (9.72)
properties of the dynamic model of the manipula-
where ˛ is a positive gain, P d and  d are the desired ac-
tor, which hold also for the constrained dynamic
celeration and velocity, respectively,  D d  , and
model (9.2). It can be easily shown that the choice of
the matrix C. q; qP / that guarantees the skew symme- Zt
try of the matrix H. P q/  2C. q; qP / in the joint space, x D . / d :
also makes the matrix ƒ. P q/  2. q; qP / skew symmet-
0
ric. This fundamental property of Lagrangian systems
is at the base of passivity-based control algorithms. The stability proof is based on the positive-definite Lya-
punov function
Rigid Environment 1
The control wrench hc can be chosen as V D sT ƒv . q/s C ˛xT K x ;
2
whose time derivative along the trajectories of (9.69),
hc D ƒ. q/Sv P r C  0 . q; qP / r C .Sv /T K .r  /
VP D  T K   ˛2 xT K x ;
C . q/ C Sf f ;
(9.67) is a definite semi-negative function. Hence,  D 0,
x D 0, and s D 0 asymptotically. Therefore, track-
where  0 . q; qP / D  Sv C ƒSP v , K is a suitable sym- ing of the desired velocity  d .t/ is ensured. Moreover,
metric and positive-definite matrix, and  r and f are the right-hand side of (9.70) remains bounded and
Part A | 9.4

properly designed control inputs. vanishes asymptotically. Hence, tracking of the de-
Substituting (9.67) into (9.2) yields sired force d .t/ can be ensured by setting f as for
the acceleration-resolved approach, according to the
choices (9.56)–(9.58).
ƒ. q/Sv sP C  0 . q; qP /s C .Sv /T K s Notice that position control can be achieved if a de-
C Sf . f  / D 0 ; (9.68) sired position rd .t/ is assigned for the vector r in (9.40),
provided that the matrices Sf and Sv are computed ac-
with sP D P r  P and s D  r  , showing that the cording to (9.32) and (9.42), and the vectors P d D rR d ,
closed-loop system remains nonlinear and coupled.  d D rPd , and x D rd r are used in (9.71) and (9.72).
 Force Control 9.4 Hybrid Force/Motion Control 215

Compliant Environment 9.4.3 Velocity-Resolved Approach

The control wrench hc can be chosen as
The acceleration-resolved approach, as well as the
hc D ƒ. q/vP r C  . q; qP /v r passivity-based approach, require modification of the
C Ks . v r  v e / C he C . q/ ; (9.73) current industrial robot controllers. As for the case of
impedance control, if the contact is sufficiently com-
where Ks is a suitable symmetric positive-definite ma- pliant, the closed-loop dynamics of a motion-controlled
trix while v r and its time derivative vP r are chosen as robot can be approximated by (9.28), corresponding to
a velocity-resolved control.
v r D v d C ˛x ; To achieve force and velocity control, according to
vP r D vP d C ˛v ; the end-effector twist decomposition (9.60), the control
input v r can be chosen as
where ˛ is the positive gain, v d and its time deriva-
tive vP d are properly designed control inputs, v D v r D Sv v C C0 Sf f ; (9.75)
v d  v e , and
with
Zt
Zt
x D v d :
v D  d .t/ C KI Πd . /  . / d ; (9.76)
0
0

Substituting (9.73) into (9.2) yields

and
ƒ. q/Ps C . q; qP /s C Ks s D 0 ; (9.74) P d .t/ C KP Œ d .t/  .t/ ;
f D (9.77)

with sP D vP r  vP e and s D v r  v e . where KI and KP are suitable symmetric and

The asymptotic stability of system (9.74) can be en- positive-definite matrix gains. Decoupling between
sured by setting velocity- and force-controlled subspaces and exponen-
tial asymptotic stability of the closed-loop system can
v d D Sv  d C C0 Sf P d ;
be proven as for the acceleration-resolved approach.
Also, since the force error has second-order dynamics,
where  d .t/ is a desired velocity trajectory and d .t/ is
an integral action can be added to (9.77) to improve the
the desired force trajectory. The stability proof is based
disturbance rejection capabilities, i. e.,
on the positive-definite Lyapunov function
P d .t/ C KP Πd .t/  .t/
f D
1
V D sT ƒ. q/s C ˛xT Ks x ; Zt
2
CKI Πd . /  . / d ; (9.78)
whose time derivative along the trajectories of (9.74),
0

VP D v T Ks v  ˛2 xT Ks x ; and the exponential asymptotic stability is guaranteed

if the matrices KP and KI are symmetric and positive
is a negative-definite function. Hence, v D 0 definite.
and x D 0, asymptotically. In the case of constant If an estimate b
C of the stiffness matrix of the en-
contact geometry and stiffness, the following equalities vironment is used in (9.75), as for the acceleration-
hold resolved approach, the exponential convergence of
to a constant d can still be ensured for both (9.77)
Part A | 9.4

P d  /
v D Sv .d  / C C0 Sf . P ;
and (9.78).
Z t
In some applications, besides the stiffness ma-
x D Sv . d  / d C C0 Sf . d  / ; trix, also the geometry of the environment is uncer-
0 tain. In these cases, a force/motion control law sim-
ilar to (9.75) can be implemented, without using the
showing that both the velocity and force tracking errors, selection matrices Sv and Sf to separate the force-
belonging to reciprocal subspaces, converge asymptoti- controlled subspace from the velocity-controlled sub-
cally to zero. space. The motion control law can be set as in (9.76),
 216 Part A Robotics Foundations

but using full velocity feedback. Also, the force con- of the force controller over the position controller en-
trol law can be set as in (9.78), but using full force sured by the presence of the integral action on the
and moment feedback. That is, both motion control force error. This approach is tested in the experiments
and force control are applied in all directions of the reported in VIDEO 687 and VIDEO 691 where the
six-dimensional (6-D) space. The resulting control, concept of task frame is also exploited. A framework
known as force control with feedforward motion or for robotic assembly, where a standard position-based
parallel force/position control guarantees force reg- robot controller is integrated with an external con-
ulation at the expense of position errors along the troller performing force-controlled skills is presented
constrained task directions, thanks to the dominance in VIDEO 692 .

9.5 Conclusions and Further Reading

This chapter has summarized the main approaches to posed in [9.3] and the implementation discussed
force control in a unifying perspective. However, there in [9.10]. Stiffness control in Cartesian coordinates was
are many aspects that have not been considered and proposed in [9.9]. Devices based on the remote center
that must be carefully taken into account when dealing of compliance were discussed in [9.25] for successful
with interaction robotic tasks. The two major paradigms mating of rigid parts. The original idea of a mechanical
of force control (impedance and hybrid force/motion impedance model used for controlling the interaction
control) are based on several simplifying assumptions between the manipulator and the environment is pre-
that are only partially satisfied in practical implemen- sented in [9.7], and a similar formulation is given
tations. In fact, the performance of a force-controlled in [9.8]. Stability of impedance control was analyzed
robotic system depends on the interaction with a chang- in [9.26] and the problems of interaction with a stiff en-
ing environment which is very difficult to model and vironment were considered in [9.27].
identify correctly. A general contact situation is far Adaptive impedance control algorithms [9.28, 29]
from being completely predictable, not only quantita- have been proposed to overcome uncertainties in the
tively, but also qualitatively: the contact configuration dynamic parameters of the robot manipulator, while ro-
can change abruptly, or be of a different type than ex- bust control schemes can be found in [9.30]. Impedance
pected. Hence, the standard performance indices used control has also been used in the hybrid force/motion
to evaluate a control system, i. e., stability, bandwidth, control framework [9.31].
accuracy, and robustness, cannot be defined by consid- A reference work on modeling 6-DOF (spatial)
ering the robotic system alone, as for the case of robot stiffness is [9.32], while the properties of spatial
motion control, but must be always referred to the par- compliance have been analyzed in detail in [9.33–
ticular contact situation at hand. Also, a classification 35]; a 6-DOF variable compliant wrist was proposed
of all these different situations is not easy, especially in in [9.36], while several studies concerning the con-
the case of dynamics environments and when the task struction of programmed compliances, optimized for
involves multiple contacts acting in parallel. specific tasks, have been proposed [9.37, 38]. The en-
Due to the inherent complexity of the force control ergy-based approach to derive a spatial impedance was
problem, a large number of research papers on this topic introduced in [9.39], using rotation matrices; various
have been published in the past three decades. A de- 6-DOF impedance control schemes based on different
scription of the state of the art of the first decade is representations of end-effector orientation, including
provided in [9.20], whereas the progress of the second the unit quaternion, can be found in [9.40]. The quater-
decade is surveyed in [9.21] and [9.22]. More recently, nion-based formulation is extended to the case of non-
two monographs on force control [9.23, 24] have ap- block-diagonal stiffness matrix in [9.41]. A rigorous
Part A | 9.5

peared. In the following, a list of references is provided, treatment of spatial impedance control in a passivity
where more details on the arguments presented in the framework can be found in [9.42].
chapter, as well as topics not covered here, can be More recently, impedance control was proposed as
found. an effective approach to enhance safety in applications
where humans and robots share the same workspace
9.5.1 Indirect Force Control and may have physical interaction. To this end, the
passive compliance of lightweight robots is combined
The concept of generalized spring and damping for with the active compliance ensured by impedance con-
force control in joint coordinates was originally pro- trol [9.43, 44].
 Force Control 9.5 Conclusionsand Further Reading 217

9.5.2 Task Specification ance problems pointed out in [9.47] were correctly
addressed, among other papers, in [9.46, 55]. Trans-
The concepts of natural and artificial constraints and position of model-based schemes from unconstrained
of compliance frame were introduced in [9.11]. These motion control to constrained cases was accomplished
ideas have been systematically developed in [9.12, 45] for adaptive control in [9.18, 56, 57] and for robust con-
within the task frame formalism. Theoretical issues on trol in [9.58].
the reciprocity of generalized force and velocity direc- Approaches designed to cope with uncertainties in
tions are discussed in [9.46, 47], while invariance in the environment geometry are the force control with
computing the generalized inverse in robotics is ad- feedforward motion scheme proposed in [9.2] and the
dressed in [9.48]. The issue of partially constrained parallel force/position control [9.19], based on the con-
tasks is considered in [9.49], where the models of cept of dominance of force control on motion control,
positive-semidefinite stiffness and compliance matrices thanks to the use of an integral action on the force er-
are developed. The problem of estimating geomet- ror. A parallel force/position regulator was developed
ric uncertainties is considered in [9.50, 51], as well in [9.59]. The integral action for removing steady-state
as the issue of linking constraint-based task specifi- force errors has traditionally been used; its stability was
cation with real-time task execution control. This ap- proven in [9.60], while robustness with respect to force
proach is generalized in [9.52], where a systematic measurement delays was investigated in [9.61, 62].
constraint-based methodology to specify complex tasks In the absence of force/torque sensors, suitable al-
is presented. gorithms can be adopted to estimate the contact force,
based, e.g., on disturbance observers [9.63], or on the
9.5.3 Hybrid Force/Motion Control error of a position control [9.64].
It has generally been recognized that force con-
Early works on force control can be found in [9.10]. trol may cause unstable behavior during contact with
The original hybrid force/positon control concept was environment. Dynamic models for explaining this phe-
introduced in [9.13], based on the natural and arti- nomenon were introduced in [9.65] and experimen-
ficial constraint task formulation [9.11]. The explicit tal investigations can be found in [9.66] and [9.67].
inclusion of the manipulator dynamic model was pre- Moreover, control schemes are usually derived on the
sented in [9.17], and a systematic approach to modeling assumption that the manipulator end-effector is in con-
the interaction with a dynamic environment was de- tact with the environment and that this contact is not
veloped in [9.53]. The constrained formulation with lost. Impact phenomena may occur and deserve careful
inverse dynamic controllers is treated in [9.14, 54] in consideration, and there is a need for global analy-
the Cartesian space as well as in [9.15] joint space. sis of control schemes including the transition from
The constrained approach was also used in [9.16] with noncontact to contact situations and vice versa, see
a controller based on linearized equations. The invari- e.g., [9.68–70].

Video-References

VIDEO 684 Recent research in impedance control

available from [Link]
VIDEO 685 Integration of force strategies and natural admittance control
available from [Link]
VIDEO 686 Experiments of spatial impedance control
available from [Link]
VIDEO 687 Compliant robot motion; Control and task specification
Part A | 9

available from [Link]

VIDEO 691 COMRADE: Compliant motion research and development environment
available from [Link]
VIDEO 692 Robotic assembly of emergency stop buttons
available from [Link]
 218 Part A Robotics Foundations

References

9.1 T.L. De Fazio, D.S. Seltzer, D.E. Whitney: The instru- 9.19 S. Chiaverini, L. Sciavicco: The parallel approach to
mented remote center of compliance, Ind. Robot force/position control of robotic manipulators, IEEE
11(4), 238–242 (1984) Trans. Robotics Autom. 9, 361–373 (1993)
9.2 J. De Schutter, H. Van Brussel: Compliant robot mo- 9.20 D.E. Whitney: Historical perspective and state of the
tion II. A control approach based on external control art in robot force control, Int. J. Robotics Res. 6(1),
loops, Int. J. Robotics Res. 7(4), 18–33 (1988) 3–14 (1987)
9.3 I. Nevins, D.E. Whitney: The force vector assembler 9.21 M. Vukobratović, Y. Nakamura: Force and contact
concept, Proc. 1 CISM-IFToMM Symp. Theory Pract. control in robotic systems, Proc. IEEE Int. Conf.
Robotics Manip., Udine (1973) Robotics Autom., Atlanta (1993)
9.4 M.T. Mason, J.K. Salisbury: Robot Hands and Me- 9.22 J. De Schutter, H. Bruyninckx, W.H. Zhu, M.W. Spong:
chanics of Manipulation (MIT Press, Cambridge 1985) Force control: A bird’s eye view. In: Control Prob-
9.5 J.Y.S. Luh, W.D. Fisher, R.P.C. Paul: Joint torque con- lems in Robotics and Automation, ed. by K.P. Vala-
trol by direct feedback for industrial robots, IEEE vanis, B. Siciliano (Springer, London 1998) pp. 1–
Trans. Autom. Control 28, 153–161 (1983) 17
9.6 G. Hirzinger, N. Sporer, A. Albu-Schäffer, M. Hähnle, 9.23 D.M. Gorinevski, A.M. Formalsky, A.Y. Schneider:
R. Krenn, A. Pascucci, R. Schedl: DLR’s torque- Force Control of Robotics Systems (CRC, Boca Raton
controlled light weight robot III – Are we reach- 1997)
ing the technological limits now?, Proc. IEEE Int. 9.24 B. Siciliano, L. Villani: Robot Force Control (Kluwer,
Conf. Robotics Autom., Washington (2002) pp. 1710– Boston 1999)
1716 9.25 D.E. Whitney: Quasi-static assembly of compliantly
9.7 N. Hogan: Impedance control: An approach to ma- supported rigid parts, ASME J. Dyn. Syst. Meas. Con-
nipulation: Parts I–III, ASME J. Dyn. Syst. Meas. Con- trol 104, 65–77 (1982)
trol 107, 1–24 (1985) 9.26 N. Hogan: On the stability of manipulators perform-
9.8 H. Kazerooni, T.B. Sheridan, P.K. Houpt: Robust ing contact tasks, IEEE J. Robotics Autom. 4, 677–686
compliant motion for manipulators. Part I: The (1988)
fundamental concepts of compliant motion, IEEE 9.27 H. Kazerooni: Contact instability of the direct drive
J. Robotics Autom. 2, 83–92 (1986) robot when constrained by a rigid environment, IEEE
9.9 J.K. Salisbury: Active stiffness control of a manipu- Trans. Autom. Control 35, 710–714 (1990)
lator in Cartesian coordinates, 19th IEEE Conf. Decis. 9.28 R. Kelly, R. Carelli, M. Amestegui, R. Ortega: Adaptive
Control, Albuquerque (1980) pp. 95–100 impedance control of robot manipulators, IASTED Int.
9.10 D.E. Whitney: Force feedback control of manipulator J. Robotics Autom. 4(3), 134–141 (1989)
fine motions, ASME J. Dyn. Syst. Meas. Control 99, 9.29 R. Colbaugh, H. Seraji, K. Glass: Direct adaptive
91–97 (1977) impedance control of robot manipulators, J. Robotics
9.11 M.T. Mason: Compliance and force control for com- Syst. 10, 217–248 (1993)
puter controlled manipulators, IEEE Trans. Syst. Man 9.30 Z. Van Lu, A.A. Goldenberg: Robust impedance con-
Cybern. 11, 418–432 (1981) trol and force regulation: Theory and experiments,
9.12 J. De Schutter, H. Van Brussel: Compliant robot mo- Int. J. Robotics Res. 14, 225–254 (1995)
tion I. A formalism for specifying compliant motion 9.31 R.J. Anderson, M.W. Spong: Hybrid impedance con-
tasks, Int. J. Robotics Res. 7(4), 3–17 (1988) trol of robotic manipulators, IEEE J. Robotics Autom.
9.13 M.H. Raibert, J.J. Craig: Hybrid position/force control 4, 549–556 (1986)
of manipulators, ASME J. Dyn. Syst. Meas. Control 103, 9.32 J. Lončarić: Normal forms of stiffness and compliance
126–133 (1981) matrices, IEEE J. Robotics Autom. 3, 567–572 (1987)
9.14 T. Yoshikawa: Dynamic hybrid position/force control 9.33 T. Patterson, H. Lipkin: Structure of robot compli-
of robot manipulators – Description of hand con- ance, ASME J. Mech. Design 115, 576–580 (1993)
straints and calculation of joint driving force, IEEE 9.34 E.D. Fasse, P.C. Breedveld: Modelling of elastically
J. Robotics Autom. 3, 386–392 (1987) coupled bodies: Part I – General theory and geo-
9.15 N.H. McClamroch, D. Wang: Feedback stabilization metric potential function method, ASME J. Dyn. Syst.
and tracking of constrained robots, IEEE Trans. Au- Meas. Control 120, 496–500 (1998)
tom. Control 33, 419–426 (1988) 9.35 E.D. Fasse, P.C. Breedveld: Modelling of elastically
9.16 J.K. Mills, A.A. Goldenberg: Force and position con- coupled bodies: Part II – Exponential and gener-
Part A | 9

trol of manipulators during constrained motion alized coordinate method, ASME J. Dyn. Syst. Meas.
tasks, IEEE Trans. Robotics Autom. 5, 30–46 (1989) Control 120, 501–506 (1998)
9.17 O. Khatib: A unified approach for motion and 9.36 R.L. Hollis, S.E. Salcudean, A.P. Allan: A six-degree-
force control of robot manipulators: The operational of-freedom magnetically levitated variable com-
space formulation, IEEE J. Robotics Autom. 3, 43–53 pliance fine-motion wrist: Design, modeling and
(1987) control, IEEE Trans. Robotics Autom. 7, 320–333
9.18 L. Villani, C. Canudas de Wit, B. Brogliato: An ex- (1991)
ponentially stable adaptive control for force and 9.37 M.A. Peshkin: Programmed compliance for error cor-
position tracking of robot manipulators, IEEE Trans. rective assembly, IEEE Trans. Robotics Autom. 6, 473–
Autom. Control 44, 798–802 (1983) 482 (1990)
 Force Control References 219

9.38 J.M. Shimmels, M.A. Peshkin: Admittance matrix de- troller design and experiment, IEEE J. Robotics Au-
sign for force-guided assembly, IEEE Trans. Robotics tom. 4, 699–705 (1988)
Autom. 8, 213–227 (1992) 9.55 J. De Schutter, D. Torfs, H. Bruyninckx, S. Dutré: In-
9.39 E.D. Fasse, J.F. Broenink: A spatial impedance con- variant hybrid force/position control of a velocity
troller for robotic manipulation, IEEE Trans. Robotics controlled robot with compliant end effector using
Autom. 13, 546–556 (1997) modal decoupling, Int. J. Robotics Res. 16(3), 340–
9.40 F. Caccavale, C. Natale, B. Siciliano, L. Villani: Six-DOF 356 (1997)
impedance control based on angle/axis representa- 9.56 R. Lozano, B. Brogliato: Adaptive hybrid force-
tions, IEEE Trans. Robotics Autom. 15, 289–300 (1999) position control for redundant manipulators, IEEE
9.41 F. Caccavale, C. Natale, B. Siciliano, L. Villani: Robot Trans. Autom. Control 37, 1501–1505 (1992)
impedance control with nondiagonal stiffness, IEEE 9.57 L.L. Whitcomb, S. Arimoto, T. Naniwa, F. Ozaki: Adap-
Trans. Autom. Control 44, 1943–1946 (1999) tive model-based hybrid control if geometrically
9.42 S. Stramigioli: Modeling and IPC Control of Interactive constrained robots, IEEE Trans. Robotics Autom. 13,
Mechanical Systems – A Coordinate Free Approach 105–116 (1997)
(Springer, London 2001) 9.58 B. Yao, S.P. Chan, D. Wang: Unified formulation of
9.43 C. Ott: Cartesian Impedance Control of Redundant variable structure control schemes for robot manip-
and Flexible-Joint Robots, Springer Tracts in Ad- ulators, IEEE Trans. Autom. Control 39, 371–376 (1992)
vanced Robotics (STAR) (Springer, Berlin, Heidelberg 9.59 S. Chiaverini, B. Siciliano, L. Villani: Force/position
2008) regulation of compliant robot manipulators, IEEE
9.44 C. Ott, A. Albu-Schäffer, A. Kugi, G. Hirzinger: On the Trans. Autom. Control 39, 647–652 (1994)
passivity based impedance control of flexible joint 9.60 J.T.-Y. Wen, S. Murphy: Stability analysis of position
robots, IEEE Trans. Robotics 24, 416–429 (2008) and force control for robot arms, IEEE Trans. Autom.
9.45 H. Bruyninckx, J. De Schutter: Specification of Force- Control 36, 365–371 (1991)
controlled actions in the task frame formalism – 9.61 R. Volpe, P. Khosla: A theoretical and experimental
A synthesis, IEEE Trans. Robotics Autom. 12, 581–589 investigation of explicit force control strategies for
(1996) manipulators, IEEE Trans. Autom. Control 38, 1634–
9.46 H. Lipkin, J. Duffy: Hybrid twist and wrench control 1650 (1993)
for a robotic manipulator, ASME J. Mech. Design 110, 9.62 L.S. Wilfinger, J.T. Wen, S.H. Murphy: Integral force
138–144 (1988) control with robustness enhancement, IEEE Control
9.47 J. Duffy: The fallacy of modern hybrid control theory Syst. Mag. 14(1), 31–40 (1994)
that is based on orthogonal complements of twist 9.63 S. Katsura, Y. Matsumoto, K. Ohnishi: Modeling of
and wrench spaces, J. Robotics Syst. 7, 139–144 (1990) force sensing and validation of disturbance observer
9.48 K.L. Doty, C. Melchiorri, C. Bonivento: A theory for force control, IEEE Trans. Ind. Electron. 54, 530–
of generalized inverses applied to robotics, Int. 538 (2007)
J. Robotics Res. 12, 1–19 (1993) 9.64 A. Stolt, M. Linderoth, A. Robertsson, R. Johansson:
9.49 T. Patterson, H. Lipkin: Duality of constrained elas- Force controlled robotic assembly without a force
tic manipulation, Proc. IEEE Conf. Robotics Autom., sensor, Proc. IEEE Int. Conf. Robotics Autom. (ICRA)
Sacramento (1991) pp. 2820–2825 (2012) pp. 1538–1543
9.50 J. De Schutter, H. Bruyninckx, S. Dutré, J. De Geeter, 9.65 S.D. Eppinger, W.P. Seering: Introduction to dynamic
J. Katupitiya, S. Demey, T. Lefebvre: Estimation first- models for robot force control, IEEE Control Syst. Mag.
order geometric parameters and monitoring contact 7(2), 48–52 (1987)
transitions during force-controlled compliant mo- 9.66 C.H. An, J.M. Hollerbach: The role of dynamic mod-
tions, Int. J. Robotics Res. 18(12), 1161–1184 (1999) els in Cartesian force control of manipulators, Int.
9.51 T. Lefebvre, H. Bruyninckx, J. De Schutter: Polyedral J. Robotics Res. 8(4), 51–72 (1989)
contact formation identification for auntonomous 9.67 R. Volpe, P. Khosla: A theoretical and experimental
compliant motion, IEEE Trans. Robotics Autom. 19, investigation of impact control for manipulators, Int.
26–41 (2007) J. Robotics Res. 12, 351–365 (1993)
9.52 J. De Schutter, T. De Laet, J. Rutgeerts, W. De- 9.68 J.K. Mills, D.M. Lokhorst: Control of robotic manipu-
cré, R. Smits, E. Aerbeliën, K. Claes, H. Bruyninckx: lators during general task execution: A discontinuous
Constraint-based task specification and estimation control approach, Int. J. Robotics Res. 12, 146–163
for sensor-based robot systems in the presence of (1993)
geometric uncertainty, Int. J. Robotics Res. 26(5), 9.69 T.-J. Tarn, Y. Wu, N. Xi, A. Isidori: Force regulation
Part A | 9

433–455 (2007) and contact transition control, IEEE Control Syst. Mag.
9.53 A. De Luca, C. Manes: Modeling robots in contact with 16(1), 32–40 (1996)
a dynamic environment, IEEE Trans. Robotics Autom. 9.70 B. Brogliato, S. Niculescu, P. Orhant: On the con-
10, 542–548 (1994) trol of finite dimensional mechanical systems with
9.54 T. Yoshikawa, T. Sugie, N. Tanaka: Dynamic hybrid unilateral constraints, IEEE Trans. Autom. Control 42,
position/force control of robot manipulators – Con- 200–215 (1997)