Interaction Control

Manipulation requires interaction

object behavior affects control of force and motion

Independent control of force and motion is not possible

object behavior relates force and motion
contact a rigid surface: kinematic constraint
move an object: dynamic constraint

Accurate control of force or motion requires detailed models of

manipulator dynamics
object dynamics
object dynamics are usually known poorly, often not at all

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 1

Object Behavior

Can object forces be treated as external (exogenous) disturbances?

the usual assumptions dont apply:
disturbance forces depend on manipulator state
forces often arent small by any reasonable measure

Can forces due to object behavior be treated as modeling uncertainties?

yes (to some extent) but the usual assumptions dont apply:
command and disturbance frequencies overlap

Example: two people shaking hands

how each person moves influences the forces evoked
disturbance forces are state-dependent
each may exert comparable forces and move at comparable speeds
command & disturbance have comparable magnitude & frequency

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 2

Alternative: control port behavior

Port behavior:
system properties and/or
behaviors seen at an
interaction port

Interaction port:
characterized by conjugate
variables that define power flow

power in
P = et f

t
e = [e1 Len ] efforts (forces)

t
f = [ f1 L f n ] flows (velocities)

Key point:

port behavior is unaffected

by contact and interaction

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 3

Impedance & Admittance

Impedance and admittance

characterize interaction
a dynamic generalization of
resistance and conductance

e( s ) 1
=
i (s ) Cs
e( s )
electrical inductor Z ( s ) =
= L(s )
i (s )

electrical capacitor Z (s ) =

Usually introduced for linear

systems but generalizes to
nonlinear systems

z& = Z s ( z ,V )
State equations
F = Z ( z ,V )
Output equations
o

state-determined representation:
t
Constraint on input & output
P
F
V
=

this form may be derived from

n
m
m

or depicted as a network model z , F ,V , P

nonlinear 1D elastic element (spring)

x& = v
f = ( x )
Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 4

Impedance & Admittance (continued)

Admittance is the causal dual of

impedance
Admittance: flow out, effort in
Impedance: effort out, flow in

Linear system: admittance is the

inverse of impedance
Nonlinear system:
causal dual is well-defined:
but may not correspond to any
impedance
inverse may not exist

Y (s ) = Z (s )
electrical capacitor
i (s )
Y (s ) =
= Cs
(
)
es
1

y& = Ys ( y , F )
V = Y ( y, F )
o

t
P
F
V
=

y n , F m ,V m , P

nonlinear 1D inertial element (mass)

p& = f
v = ( p )
Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 5

Impedance as dynamic stiffness

Impedance is also loosely

defined as a dynamic
generalization of stiffness
effort out, displacement in

Most useful for mechanical

systems

z& = Z s ( z , X )
F = Z (z, X )
o

t
dW
F
dX
=

z n , F m , X m , P

displacement (or generalized

position) plays a key role

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 6

Interaction control: causal considerations

Whats the best input/output form for the manipulator?

The set of objects likely to be manipulated includes
inertias
minimal model of most movable objects
kinematic constraints
simplest description of surface contact

Causal considerations:
inertias prefer admittance causality
constraints require admittance causality
compatible manipulator behavior should be an impedance

An ideal controller should make the manipulator behave as an

impedance
Hence impedance control
Hogan 1979, 1980, 1985, etc.

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 7

Robot Impedance Control

Works well for interaction tasks:

Automotive assembly
(Case Western Reserve
University, US)
Food packaging
(Technical University Delft,
NL)
Hazardous material handling
(Oak Ridge National Labs,
US)
Automated excavation
(University of Sydney,
Australia)
and many more

Mod. Sim. Dyn. Sys.

Facilitates multi-robot / multi-limb

coordination
Schneider et al., Stanford

Enables physical cooperation of

robots and humans

Interaction Control

Kosuge et al., Japan

Hogan et al., MIT

Neville Hogan page 8

OSCAR the robot

Photograph removed due to copyright restrictions.

E.D.Fasse & J.F.Broenink, U. Twente, NL

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 9

Network modeling perspective on interaction control

Port concept
control interaction port behavior
port behavior is unaffected by contact and interaction

Causal analysis
impedance and admittance characterize interaction
object is likely an admittance
control manipulator impedance

Model structure
structure is important
power sources are commonly modeled as equivalent networks
Thvenin equivalent
Norton equivalent

Can equivalent network structure be applied to interaction control?

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 10

Equivalent networks

Initially applied to networks of static linear elements

Sources & linear resistors
Thvenin equivalent network
M. L. Thvenin, Sur un nouveau thorme dlectricit dynamique.
Acadmie des Sciences, Comptes Rendus 1883, 97:159-161

Thvenin equivalent sourcepower supply or transfer

Thvenin equivalent impedanceinteraction
Connectionseries / common current / 1-junction
Norton equivalent network is the causal dual form

Subsequently applied to networks of dynamic linear elements

Sources & (linear) resistors, capacitors, inductors

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 11

Nonlinear equivalent networks

Can equivalent networks be defined for nonlinear systems?

Nonlinear impedance and admittance can be defined as above
Thvenin & Norton sources can also be defined
Hogan, N. (1985) Impedance Control: An Approach to Manipulation.
ASME J. Dynamic Systems Measurement & Control, Vol. 107, pp. 1-24.

However
In general the junction structure cannot

In other words:
separating the pieces is always possible
re-assembling them by superposition is not

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 12

Nonlinear equivalent network for interaction control

V0 = V0 : {c}

One way to preserve the

junction structure:

V = V0 V

specify an equivalent network

structure in the (desired)
interaction behavior
provides key superposition
properties

virtual trajectory

network junction structure (0 junction)

z& = Zs (z, V ) : {c}
nodic impedance
(
)
{
}
:
c
F = Z o z, V

: {c} denotes modulation by control inputs

Specifically:
nodic desired impedance
does not require inertial
reference frame
virtual trajectory
virtual as it need not be a
realizable trajectory

Mod. Sim. Dyn. Sys.

Norton equivalent network

Interaction Control

Neville Hogan page 13

Virtual trajectory
Nodic impedance:

Vo

V
k
m
b

virtual trajectory

interaction port

Vo

Virtual trajectory:
like a motion controllers
reference or nominal trajectory
but no assumption that
dynamics are fast compared to
motion
virtual because it need not be
realizable
e.g., need not be confined
to manipulators workspace

V
k
m
F

nodic impedance

Vo:Sf

F
0

object

V
1

Defines desired interaction

dynamics
Nodic because input velocity is
defined relative to a virtual
trajectory

I :m

interaction
port

virtual
trajectory

1
1/k: C

R:b

nodic impedance

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 14

Superposition of impedance forces

Minimal object model is an

inertia
it responds to the sum of input
forces
in network terms: it comes with
an associated 1-junction

This guarantees linear

summation of component
impedances
even if the component
impedances are nonlinear

V1 = Vo1 V

z& 1 = Z s1 (z1 , V1 )

F1 = Z o1 (z1 , V1 )

V2 = Vo 2 V

z& 2 = Z s 2 (z 2 , V2 )

& = m 1 (F1 + F2 + F3 )
V

F2 = Z o 2 (z 2 , V2 )

V3 = Vo3 V

z& 3 = Z s 3 (z 3 , V3 )

F3 = Z o3 (z 3 , V3 )
Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 15

One application: collision avoidance

Impedance control also enables non-contact (virtual) interaction

Impedance component to acquire target:
Attractive force field (potential valley)
Impedance component to prevent unwanted collision:
Repulsive force-fields (potential hills)
One per object (or part thereof)
Total impedance is the sum of these components
Simultaneously acquires target while preventing collisions
Works for moving objects and targets
Update their location by feedback to the (nonlinear) controller
Computationally simple
Initial implementation used 8-bit Z80 processors
Andrews, J. R. and Hogan, N. (1983) Impedance Control as a
Andrews & Hogan, 1983
Framework for Implementing Obstacle Avoidance in a Manipulator,
pp. 243-251 in D. Hardt and W.J. Book, (eds.), Control of
Manufacturing Processes and Robotic Systems, ASME.

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 16

High-speed collision avoidance

Static protective (repulsive) fields must extend beyond object

boundaries
may slow the robot unnecessarily
may occlude physically feasible paths
especially problematical if robot links are protected

Solution: time-varying impedance components

protective (repulsive) fields grow as robot speeds up, shrink as it slows
down
Fields shaped to yield maximum acceleration or deceleration
Newman, W. S. and Hogan, N. (1987) High Speed Robot Control
Newman & Hogan, 1987
and Obstacle Avoidance Using Dynamic Potential Functions, proc.
IEEE Int. Conf. Robotics & Automation, Vol. 1, pp. 14-24.

See also extensive work by Khatib et al., Stanford

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 17

Impedance Control Implementation

Controlling robot impedance is an ideal

like most control system goals it may be difficult to attain

How do you control impedance or admittance?

One primitive but highly successful approach:
Design low-impedance hardware
Low-friction mechanism
Kinematic chain of rigid links

Torque-controlled actuators
e.g., permanent-magnet DC motors
high-bandwidth current-controlled amplifiers

Use feedback to increase output impedance

(Nonlinear) position and velocity feedback control

Simple impedance control

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 18

Robot Model

Effort-driven inertia

& + C(, ) + G ( ) = motor + interaction

I ( )
: generalized coordinates, joint angles, configuration
variables
: generalized velocities, joint angular velocities

Linkage kinematics transform

interaction forces to interaction
torques
X = L( )
& = (L )& = J ( )
V=X

: generalized forces, joint torques

interaction = J ( )t Finteraction

I: configuration-dependent inertia

X: interaction port (end-point) position

C: inertial coupling (Coriolis & centrifugal

accelerations)

V: interaction port (end-point) velocity

G: potential forces (gravitational torques)

Finteraction: interaction port force

L: mechanism kinematic equations
J: mechanism Jacobian

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 19

Simple Impedance Control

Fimpedance = (Xo X ) + B(Vo V )

Target end-point behavior

Norton equivalent network with
elastic and viscous impedance,
possibly nonlinear

Express as equivalent (jointspace) configuration-space

behavior
use kinematic transformations

This defines a position-andvelocity-feedback controller

A (non-linear) variant of PD
(proportional+derivative)
control

that will implement the target

behavior

Mod. Sim. Dyn. Sys.

Xo: virtual position

Vo: virtual velocity
K: displacement-dependent (elastic) force function
B: velocity-dependent force function
t
motor = J ( ) Fimpedance

motor = J ( )t ( (Xo L( )) + B(Vo J ( ) ))

Dynamics of controller impedance coupled
to mechanism inertia with interaction port:

& + C(, ) + G ( ) =
I( )

J ( )t ( (Xo L( )) + B(Vo J ( ) ))

+ J ( )t Finteraction

Interaction Control

Neville Hogan page 20

Mechanism singularities

Impedance control also facilitates interaction with the

robots own mechanics
Compare with motion control:

Position control maps desired end-point trajectory onto

configuration space (joint space)
Requires inverse kinematic equations
Ill-defined, no general algebraic solution exists

X = L( )
desired = L1 (X desired )

one end-point position usually corresponds to many

configurations
some end-point positions may not be reachable

Resolved-rate motion control uses inverse Jacobian

A typical motion controller wont work at or near these

singular configurations

V = J ( )
Locally linear approach, will find a solution if one exists
desired = J ( )1 Vdesired
At some configurations Jacobian becomes singular
Motion is not possible in one or more directions

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 21

Mechanism junction structure

Mechanism kinematics relate

configuration space {} to
workspace {X}
In network terms this defines a
multiport modulated
transformer
Hence power conjugate
variables are well-defined in
opposite directions

Mod. Sim. Dyn. Sys.

Generalized coordinates
uniquely define mechanism
configuration
By definition

Hence the following maps are

always well-defined
generalized coordinates
(configuration space) to endpoint coordinates (workspace)
generalized velocities to
workspace velocity
workspace force to generalized
force
workspace momentum to
generalized momentum

Interaction Control

Neville Hogan page 22

Control at mechanism singularities

Simple impedance control law was derived by transforming desired

behavior
Norton equivalent network in workspace coordinates

from workspace to configuration (joint) space

All of the required transformations are guaranteed well-defined at all
configurations
X
V
F

motor = J ( )t ( (Xo L( )) + B(Vo J ( ) ))

Hence the simple impedance controller can operate near, at and

through mechanism singularities

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 23

Generalized coordinates

Aside:
Identification of generalized coordinates requires care
Independently variable
Uniquely define mechanism configuration
Not themselves unique
Actuator coordinates are often suitable, but not always
Example: Stewart platform
Identification of generalized forces also requires care
Power conjugates to generalized velocities
or
dW = td
P = t
Actuator forces are often suitable, not always

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 24

Inverse kinematics

Generally a tough computational problem

Modeling & simulation afford simple, effective solutions
Assume a simple impedance controller
Apply it to a simulated mechanism with simplified dynamics
Guaranteed convergence properties Hogan, N. (1984) Some Computational Problems
Hogan 1984
Slotine &Yoerger 1987

Simplified by Impedance Control, proc. ASME Conf. on

Computers in Engineering, pp. 203-209.
Slotine, J.-J.E., Yoerger, D.R. (1987) A Rule-Based
Inverse Kinematics Algorithm for Redundant
Manipulators Int. J. Robotics & Automation 2(2):86-89

Same approach works for redundant mechanisms

Redundant: more generalized coordinates than workspace coordinates

Inverse kinematics is fundamentally ill-posed
Rate control based on Moore-Penrose pseudo-inverse suffers drift
Proper analysis of effective stiffness eliminates drift
Mussa-Ivaldi & Hogan 1991

Mod. Sim. Dyn. Sys.

Mussa-Ivaldi, F. A. and Hogan, N. (1991) Integrable

Solutions of Kinematic Redundancy via Impedance
Control. Int. J. Robotics Research, 10(5):481-491

Interaction Control

Neville Hogan page 25

Intrinsically variable impedance

Feedback control of impedance suffers inevitable imperfections

parasitic sensor & actuator dynamics
communication & computation delays

Alternative: control impedance using intrinsic properties of the

actuators and/or mechanism
Stiffness
Damping
Inertia

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 26

Intrinsically variable stiffness

1

Engineering approaches

Moving-core solenoid

Moving-core solenoid
Separately-excited DC machine
Fasse et al. 1994
Variable-pressure air cylinder
Pneumatic tension actuator
McKibben muscle
and many more

0.5

0
-2

-1.5

-1

-0.5

0.5

1.5

length
-0.5

force

-1

Mammalian muscle
antagonist co-contraction increases
stiffness & damping
complex underlying physics
see 2.183
increased stiffness requires
increased force

Mod. Sim. Dyn. Sys.

Interaction Control

Fasse, E. D., Hogan, N., Gomez, S. R., and Mehta, N.

R. (1994) A Novel Variable Mechanical-Impedance
Electromechanical Actuator. Proc. Symp. Haptic
Interfaces for Virtual Environment and Teleoperator
Systems, ASME DSC-Vol. 55-1, pp. 311-318.

Neville Hogan page 27

Opposing actuators at a joint

Assume
constant moment arms
linear force-length relation
(grossly) simplified model of
antagonist muscles about a
joint

f g = k g lg
t = rg f g rn f n = rg k g (lgo rg q ) rn k n (lno + rn q )

q: joint angle; t: torque; K: joint stiffness

subscripts: g: agonist; n: antagonist, o: virtual

Equivalent behavior:
Opposing torques subtract
Opposing impedances add

t = (rg k g lgo rn k n lno ) rg2 k g + rn2 k n q

t = K (q o q )

Joint stiffness positive if actuator

stiffness positive

Mod. Sim. Dyn. Sys.

l n = l no + rn q
f n = k n ln

f: force; l: length; k: actuator stiffness

lg = lgo rg q

q o = (rg k g lgo rn k n lno )

K = rg2 k g + rn2 k n

Interaction Control

Neville Hogan page 28

Configuration-dependent moment arms

Connection of linear actuators

usually makes moment arm
vary with configuration
lg = lg (q )

lg q = rg (q ) < 0

f g = k g (lgo lg (q ))

l n = ln (q )

ln q = rn (q ) > 0

f n = k n (l no l n (q ))

Mod. Sim. Dyn. Sys.

Joint stiffness, K:
Second term always positive
First term may be negative
r
r
K = g f g + n f n + rg2 k g + rn2 k n
q
q

More typical:
change signs on
the transformers

t = rg (q )f g (q ) rn (q )f n (q )
r
f
t
r
f
= g f g rg g n f n rn n
q
q
q q
q
Interaction Control

Neville Hogan page 29

This is the tent-pole effect

Consequences of configurationdependent moment arms:

Opposing ideal (zero-impedance)
tension actuators
agonist moment grows with angle,
antagonist moment declines
always unstable

Constant-stiffness actuators
stable only for limited tension

Mammalian muscle:
stiffness is proportional to tension
good approximation of complex
behavior
can be stable for all tension

Mod. Sim. Dyn. Sys.

Take-home messages:
Kinematics matters
Kinematic stiffness may
dominate

Impedance matters
Zero output impedance may be
highly undesirable

Interaction Control

Neville Hogan page 30

Intrinsically variable inertia

Inertia is difficult to modulate via feedback but mechanism inertia is a

strong function of configuration
Use excess degrees of freedom to modulate inertia
e.g., compare contact with the fist or the fingertips

Consider the apparent (translational) inertia at the tip of a 3-link openchain planar mechanism
Use mechanism transformation properties

Translational inertia is usually characterized by

Generalized (configuration space) inertia is
Jacobian:

v = J ( )
= J ( )t p

Corresponding tip (workspace) inertia:

p = Mv
= I ( )

p = J ()- t I () J ( )-1 v
M tip = J ()- t I ()J ( )-1

Snag: J() is not squareinverse J()-1 does not exist

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 31

Causal analysis

v = M 1p
= I( )1

Inertia is an admittance
prefers integral causality

Transform inverse configuration-space inertia

Corresponding tip (workspace) inertia
This transformation is always well-defined

v = J ( )I ( )1 J ( )t p
M tip1 = J ( )I ( )1 J ( )t

Does I()-1always exist?

consider how we constructed I() from individual link inertias
I() must be symmetric positive definite, hence its inverse exists

Does Mtip-1 always exist?

yes, but sometimes it loses rank
inverse mass goes to zero in some directionscant move that way
causal argument: input force can always be applied
mechanism will figure out whether & how to move

Mod. Sim. Dyn. Sys.

Interaction Control

Neville Hogan page 32