Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 55-37 (2022) 369–375
A
A Failure
Failure Identification
Identification and
and Recovery
Recovery
A
A Failure
Failure
Framework Identification
Identification
for a Planar and
and Recovery
Recovery
Reconfigurable
Framework
A Failure for a Planar Reconfigurable
Identification and Recovery
Framework
Framework for
for a
a Planar
Planar Reconfigurable
⋆
Reconfigurable
Cable
Framework Driven
Cable Driven
for a Parallel
Parallel
Planar Robot ⋆⋆⋆
Robot
Reconfigurable
Cable Driven
Cable Driven
Driven Parallel
Parallel Robot
Robot ⋆∗
Cable
Adhiti Raman ∗
∗ Parallel
∗ Robot
∗ Ian Walker ∗ Venkat Krovi ∗
∗
Adhiti Raman Ian Walker Venkat Krovi ∗
∗ ∗
Adhiti Raman ∗ Ian Walker
Matthias
∗
Matthias Schmid
Schmid ∗ ∗
∗
∗ Venkat
∗ Krovi ∗∗∗
Adhiti Raman ∗ Ian Walker ∗ Venkat ∗
∗ Krovi ∗
Adhiti Raman Ian Walker
Matthias
Matthias SchmidVenkat
Schmid ∗ Krovi
∗ ∗
∗
∗ Clemson University, MatthiasClemson, Schmid SC 29607, USA (e-mail:
∗ Clemson University, Clemson, SC 29607, USA (e-mail:
∗
∗ Clemsonadhitir,iwalker,vkrovi,schmidm@clemson.edu).
University,
∗ Clemson University, Clemson, SC 29607, USA (e-mail:
Clemson,
adhitir,iwalker,vkrovi,schmidm@clemson.edu). SC 29607, USA (e-mail:
adhitir,iwalker,vkrovi,schmidm@clemson.edu).
Clemson University, Clemson, SC 29607, USA (e-mail:
adhitir,iwalker,vkrovi,schmidm@clemson.edu).
adhitir,iwalker,vkrovi,schmidm@clemson.edu).
Abstract:
Abstract: In In cable
cable driven
driven parallel
parallel robots
robots (CDPRs),
(CDPRs), aa single single cable
cable malfunction
malfunction usually usually induces
induces
Abstract:
complete
complete In
failure cableof driven
the parallel
entire robot. robots (CDPRs),
However, the a
lostsingle
static cable malfunction
workspace (due usually
(due to
to failure)induces
can
Abstract:failure In cable of the
drivenentire robot.
parallel However,
robots (CDPRs), the lost static
a single workspace
cable malfunction failure)
usually can
induces
complete
often
Abstract:
often be
be failure
recovered
In
recoveredcableof the
through
driven
through entire robot.
reconfiguration
parallel However,
robots
reconfiguration of
(CDPRs),
of the
the
the lost
cable
a
cable static
attachment
single workspace
cable
attachment points
malfunction
points (due
on
on to
the failure)
frame.
usually
the frame. can
This
induces
This
complete failure of the entire robot. However, the lost static workspace (due to failure) can
capability
complete
often be
capability is introduced
failure
recovered of
is introduced the
through by
entireadding
robot. kinematic
reconfiguration
by adding However, of redundancies
the
the
kinematicofredundancies lost
cable to
static
attachmentthe robot
workspace
to the robot pointsin the
(due
in theon form
to
the of moving
failure)
frame.
formframe. can
This
of moving
often be recovered through reconfiguration the cable attachment points on the This
linear
often sliders
be
capability
linear slidersis that
recovered are
introduced manipulated
through by adding in a
reconfiguration real-time
kinematic of redundancy
the cable
redundancies resolution
attachment
to the controller.
points
robot in on
the theThe
form presented
frame.
of This
moving
capability is that are manipulated
introduced by addinginkinematic
a real-time redundancy to
redundancies resolution
the robot controller.
in the form The ofpresented
moving
linear
work sliders
combines
capability
work is that arecontroller
this
introduced manipulated within
with
by adding ana online
real-time
kinematic redundancy
failure detection
redundancies resolution
to framework
the robot controller.
to
in develop
to the Theaaof
form presented
complete
moving
linearcombines
sliders that this
arecontroller
manipulated inana online
real-time failure detection
redundancy framework
resolution develop
controller. complete
The presented
work tolerant
fault
linear
fault combines
sliders
tolerant that this
control
are
control controller
scheme
manipulated
scheme with
for
for inana online
automatic
automatic real-time failure
task
task detection
recovery.
redundancy
recovery. This framework
solution
resolution
This solution to develop
provides
controller.
provides The a presented
complete
robustness
robustness by
by
work combines this controller with an online failure detection framework to develop a complete
combining
work
fault combines
tolerant
combining pose
pose estimation
this
control controller
estimationscheme of the
with
for
of for end-effector
an
automatic
the automatic online
end-effector taskwith
failure
with the
recovery. failure
detection
the failure This framework
solution
detection through
to develop
provides the application
a complete
robustness by
fault tolerant control scheme task recovery. Thisdetection
solutionthrough
providesthe application
robustness by
of an
fault
combining
of Interactive
tolerant pose
an Interactive Multiple
control
estimationscheme
Multiple of Model
of for
the
Model (IMM)
automatic
end-effector algorithm
(IMM) algorithm taskwith therelying
recovery. This
failure only on
solution
detection end-effector
provides
through information.
robustness
the application by
combining pose estimation the end-effector with therelying failure only on end-effector
detection through theinformation.
application
of
The an
combining
The Interactive
failure and
pose Multiple
pose estimation
estimation Model
of the (IMM)
scheme
end-effectoris algorithm
then tied
with relying
into
the the
failure only
redundancyon
detection end-effector
resolution
through theinformation.
approach
approach to
application
of anfailure and pose
Interactive estimation
Multiple Model scheme
(IMM) is then
algorithmtied into the redundancy
relying resolution information.
only on end-effector to
produce
of
The an a
failure seamless
Interactive
and pose automatic
Multiple
estimation task
Model scheme(trajectory)
(IMM) is algorithm
then recovery
tied relying
into approach
the only
redundancyfor
on cable
end-effectorfailures.
resolution information.
failures.approach
produce
The failure a seamless
and poseautomatic
estimationtask scheme(trajectory)
is then tied recovery
into the approach
redundancyfor cableresolution approach to to
The
producefailure
Copyright
produce a
a© and
2022pose
seamless
seamless estimation
Theautomatic
Authors. This
automatic scheme
task
task is then
is (trajectory)
an open
(trajectory) accesstied intounder
article
recovery
recovery the redundancy
the CCfor
approach
approach for resolution
BY-NC-ND
cable
cable licenseapproach to
failures.
failures.
produce
Keywords:
a seamless
(https://creativecommons.org
Keywords: Robotics, automatic
Robotics, identification /licenses/by-nc-nd/4.0/)
task and
identification
(trajectory)
and fault
fault detection,
recovery approach
detection, parallel
parallel robots,
for cable failures. and
robots, identification
identification and
Keywords:
control
control Robotics,
methods,
methods, fault identification
accommodation
faultidentification
accommodation and fault
and detection,
reconfiguration parallel robots,
strategies. identification and
Keywords: Robotics, and andfaultreconfiguration
detection, parallel strategies.
robots, identification and
control methods, fault accommodation and reconfiguration strategies. identification and
Keywords:
control Robotics,
methods, fault identification
accommodation and fault
and detection,
reconfiguration parallel robots,
strategies.
control1.methods, fault accommodation and reconfiguration strategies. at the end-effector, this introduces
1. INTRODUCTION
INTRODUCTION exert
exert positive
positive tension
tension at the end-effector, this introduces
1. INTRODUCTION aexert
layer
aexert positive
of
layerpositive tensionto
complexity
of complexity atmore
the end-effector,
traditional
traditional forms this introduces
of
of control.
1. INTRODUCTION tensiontoatmore the end-effector, forms control.
this introduces
Cable driven 1.
parallel INTRODUCTION
robots (CDPRs) are lightweight The
exert
a
The flexibility
layer positive
of of
complexity
flexibility of the
tension
the cable
to at
cable introduces
the
more unmodeled
end-effector,
traditional
introduces this
forms
unmodeled ofdynamics
introduces
control.
dynamics
Cable driven parallel robots (CDPRs) are lightweight a layer of complexity to more traditional forms of control.
Cable
mechanismsdriven in parallel
which robots
cables (CDPRs)
replace are
traditional lightweight
rigid link aThe
layer
and
and
The of
actuator
actuator complexity
flexibility
flexibility of the
faults
faults
of the to
cable
which
which
cable more can traditional
introduces
can lead
lead
introduces to forms
unmodeled
to large
large
unmodeled of control.
dynamics
uncontrolled
uncontrolled
dynamics
mechanisms in which cables replace
Cable driven parallel robots (CDPRs) are lightweight and traditional rigid link actuator faults which can lead to large uncontrolled
Cable
mechanisms
actuators. driven in
They parallel
which
can be robots
cables
designed (CDPRs)
replace
and are
traditional
structured to provide
lightweight
rigid link The flexibility
tensions in the of the cable
cables and introduces
therefore unmodeled
can cause dynamics
failures.
actuators. They
mechanisms can becables
in which designed and structured
replace traditional to link tensions
provide
rigid and actuator
tensions
Here,
in the
in the
kinematic
cables
faults which
cables
redundancy
andcan
andcan
therefore
lead to can
therefore
may be
largecause
to can
largecause
failures.
uncontrolled
failures.
mechanisms
actuators. They
manipulability
manipulability
actuators. in which
They can be
over
over
can be cables
designed
large
large replace
workspaces,
designed and traditional
and structured
workspaces, structured
with
with aa high rigid
to
high
to link and
provide
ratio
ratio
provide Here, actuator
tensionskinematic
in the faults which
redundancy
cables and may be employed
lead
therefore employed
can
to
to overcome
uncontrolled
cause overcome
failures.
manipulability
of stiffness
actuators. to over
mass, large
thus workspaces,
proving with
useful in a high
applicationratio cable
tensions
Here, or actuator
in
kinematic the failures
cables
redundancy and by utilizing
therefore
may be geometric
can
employed cause
to reconfig-
failures.
overcome
of stiffnessThey
manipulability can be
to mass,
over designed
thus
large proving and useful
workspaces, structured to provide
withina application
high ratio cable uration
or actuator
Here, kinematic
to actively
failures bymay
redundancy
rearrange
utilizing
the
geometric
be employed
attachment
reconfig-
to overcome
points in
of stiffness
domains such
manipulability to asmass,
painting
over thus
large proving
(Gagliardini
workspaces, useful
et
with in
al. a application
(2015)),
high in-
ratio Here,
cable kinematic
or actuator redundancy
failures by may be
utilizing employed
geometric to overcome
reconfig-
domains
of stiffnesssuchto as painting
mass, thus (Gagliardini
proving useful et al.
in (2015)),
application in- uration cable or to
uration
the to
structure.
actively
actuator
actively
This
rearrange
failures
rearrange
ensures
the attachment
by utilizing
the
that
geometric points
attachment
the static points
traversable
in
reconfig-
in
of stiffness
domains
spection suchto
(Izard asmass,
etpainting
al. thus
(2013)), proving
(Gagliardini useful
warehousing, et in
al.
and application
(2015)),
manufac- in- cable or actuator failures by utilizing geometric reconfig-
spection (Izard
domains such asetpainting
al. (2013)), warehousing,
(Gagliardini et al.and manufac-
(2015)), in- the the
structure.
uration
workspace
uration
to actively
structure.
to lost
This rearrange
This
through
actively
ensures that
ensures
cable
rearrange
the the
that
failure
the theis
static traversable
attachment
static
quickly
attachment
points in
traversable
recovered.
points in
domains
turing,
turing, such aset
spectionincluding
(Izard
including small-scale
etpainting
al. (2013)),
(2013)),
small-scale and
(Gagliardini
and large-scale
warehousing, et al.
large-scale and3D
3D printing
(2015)),
manufac-
printing in- workspacethe structure. lost through
This ensures cable thatfailuretheis quickly recovered.
static traversable
spection (Izard al. warehousing, and manufac- Furthermore,
the structure.
workspace lost an online-redundancy
This
through ensures
cable that
failuretheresolution
is static
quickly algorithm
traversable
recovered.
turing,
(Chesser including
et
spectionincluding
(Izardal.
et al. et small-scale
(2022); and
Jamshidifar
al. (2013)), large-scale
warehousing,et al. 3D
(2015);
et al. and printing
Izard
(Chesser
turing, (2022);
small-scaleJamshidifar
and large-scale 3Dmanufac-
(2015); Izard Furthermore,
printing workspace lostan
workspace
Furthermore, lostthe
an
online-redundancy
through
through
cable failure resolution
cable
online-redundancy failuretravels
is quickly
resolution
algorithm
is quickly recovered.
recovered.
algorithm
et
et al.
turing,
(Chesser
al. (2017)).
including
et
(2017)). al. Traditional
small-scale
(2022);
Traditional CDPRs
and
Jamshidifar
CDPRs have
et
have fixed
large-scale
al. 3D workspace
(2015);
fixed printing
workspaceIzard ensures
ensures
Furthermore, that
that the
an static
static workspace
workspace
online-redundancy travels with
with
resolution the
the end-
end-
algorithm
(Chesser et al. (2022); Jamshidifar et al. (2015); Izard effector. ensures that the static workspace travels
(Chesser
et al.
domains
domains
et et
(2017)).
and al. (2022);
Traditional
generally
and generally
al. (2017)). Traditional Jamshidifar
CDPRs
invariant
invariant
CDPRs et
have
maps
maps al.
of (2015);
fixed
haveoffixed workspace
workspace
workspace
workspaceIzard
and
and effector.
Furthermore,
ensures that
The basic
The basican idea is depicted
online-redundancy
idea isworkspace
the static depicted in in Fig. 1.with the
Fig.
travels
with
1.
resolution end-
algorithm
the end-
et al. (2017)).
domains
performance Traditional
and quality
generally within those
CDPRs
invariant domains.
haveof
maps Here,
offixed the
workspace
workspace ad-
and ensures
effector.
effector. that
The
Theis the
basic
basic static
idea
idea workspace
is depicted
is depicted travels
in Fig.
inthat
Fig.aimswith
1.
1. at the end-
performance
domains and quality
generally within those maps
invariant domains. Here,
workspace the and ad- This This study
study is part
part of
of aa larger
larger effort
effort that aims at advanc-
advanc-
performance
dition
domains of quality
modularity within
to the those
design domains.
in the formHere, of the
geomet- ad- effector. The basic idea is depicted in Fig. 1.
dition of and
performance generally
modularity
quality to invariantthose maps
the design
within in theofform
domains. workspace
Here, the and
of geomet- ad- ing This
ing study
the is part of aof
development larger effort thatrobot
aa cooperative aims (cobot)
at advanc-for
ric reconfigurability
performance
dition of quality
modularity offers
within
to the the
thoseadvantageous
design domains.
in the form flexibility
Here, of the
geomet- to
ad- Thisthestudy development
is part of aoflarger cooperative
effort thatrobotaims (cobot)
at advanc-for
ric reconfigurability offers the advantageous
dition of modularity to the design in the form of geomet- assisting flexibility to assisting
This
ing study
the concrete
is part
development delivery
of
concrete delivery a
of larger
a in tasks
effort
cooperative
in tasks suchsuch
that as
aims
robot 3D
as 3D at concrete
advanc-
(cobot) for
concrete
dition
ric of modularity
reconfigurability
improve performance to
offers
at the design
the
the in the
advantageous
end-effector form
or of geomet-
flexibility
to move theto ing the development of a cooperative robot (cobot) for
improve
ric performanceoffers
reconfigurability at the theend-effector
advantageous or to move the
flexibility to ing printing
the
assisting
printing or human
development
concrete directed
of
delivery a concrete
cooperative
in tasks suchapplications.
robot
as 3D(cobot) This
for
concrete
ric reconfigurability
improve
robot into performance
previously inaccessible
offers
at the
the workspace
advantageous
end-effector ordomains.
flexibility
to move This
theto assisting or humandelivery
concrete directedin concrete
tasks such applications.
as 3D concrete This
robot intoperformance
improve previously inaccessible
at the end-effectorworkspaceordomains.
to moveThis the assisting cobot
printing is expected
concrete
or
cobot is expected human to consist
delivery
directed
to directed of
in two
taskssubsystems:
concrete
consist ofconcrete such as 3D (i)
applications.
two subsystems: a cable
concrete
This
(i) a cable
robot
has
has into previously
been
improve demonstrated
intoperformance
been demonstrated inaccessible
at in thenumerous
in workspace
end-effector
numerous recentdomains.
toapproaches
ordomains.
recent moveThis
approaches the printing
This cobot
printing
or human
is expected
or human to directed
consist of two subsystems:
applications. This
(i) a cable
robot
robot
has into
been
(Rasheed
previously
previously
demonstrated
(2019); Zhou
inaccessible
inaccessible
et in
al. numerous
(2014);
workspace
workspacerecent
Raman domains.
et approaches
al. This
(2020); cobot is expected to consist ofconcrete applications.
two subsystems: This
(i) a cable
(Rasheed
has been (2019); Zhou etinal.numerous
demonstrated (2014); Raman recentetapproaches
al. (2020); cobot is expected to consist of two subsystems: (i) a cable
has
Serianibeenet
(Rasheed
Seriani
(Rasheed demonstrated
(2019);
al. Zhou et
al. (2016)).
et(2019);(2016)).
Zhou Ininal.
et
In al. numerous
this
this(2014);
study,
study,Raman
(2014); recent
Raman
we
we etapproaches
al. (2020);
demonstrate
demonstrate
et al. (2020);aa
(Rasheed
Seriani
new et(2019);
advantage
new advantage
Seriani al. of Zhou
(2016)). et
In
incorporating
of incorporating
et al. (2016)). al.
this(2014);
study,
modularity
In this modularity Raman
we
study, we by et al.
demonstrate
by utilizing
utilizing
demonstrate (2020);the
the aa
new advantage
offered
Seriani
offered kinematic
et al.
kinematic of incorporating
redundancies
(2016)). In
redundancies this modularity
for failure
study,
for failurewe by utilizing
tolerant
demonstrate
tolerant control.
control.the a
new advantage of incorporating modularity by utilizing the
new
offeredadvantage
offered kinematic
kinematic of incorporating
redundancies for
redundancies modularity
for by utilizing
failure tolerant
failure tolerant control.
control.the
Maintaining
Maintaining
offered kinematic robot operation
operation despite
robotredundancies despite
for failure actuator
tolerantfailure
actuator failure
control. is
is
Maintaining
a
a rich
rich field
field robot
of
of study,
study,operation
especially
especially despite
in
in actuator and
aerospace
aerospace failure
and space
space is
Maintaining robot operation despite actuator failure is
Maintaining
a rich
robotics
arobotics field
where
where
rich field robot
of study,operation
mechanical
of mechanical especially
study, especially and despite
in
controller
and controller actuator
aerospace
redundancies
redundancies
in aerospace failure
and
and spacespace
are
areis
a rich
robotics
design field
where of
critical.
design critical.
robotics study,
mechanical
Despite especially
Despite cable
where mechanical and
cable in
failuresaerospace
controller being
failures being
and controller aand
redundancies common
a common
redundancies space
are
are
design
problem critical.
roboticscritical.
where
in
in CDPRs, Despite
mechanical
however, cable failures on
andresearch
controller being a common
redundancies are
problem
design CDPRs, however,
Despite cable failures on
research failure
failure
being tolerant
tolerant
a common (a) (b)
problemcritical.
control
design
control in
ininthis
CDPRs,
this domain however,
Despite
domain remains
cable
remains research
sparse.
failures
sparse. on
As failure
cables
being
As cables a tolerant
can
common
can only
only (a) (b)
problem in CDPRs, however, research on failure tolerant (a) (b)
problem
control in
control ininthis
CDPRs,
this domain
domain however,
remainsresearch
remains sparse. As
sparse. on cables
As failure can
cables tolerant
can only
only (a) (b)
⋆
⋆control in this
This work domain remains
is supported in part by sparse. As cables
the U.S. National canScience Fig. 1. (a) Cable(a) failure
only Fig. 1. (a) Cable failure (b) Reconfigurability (b) Reconfigurability
(b) allows
allows the
the
This work is supported in part by the U.S. National Science trajectory tracking to continue
⋆
Foundation
⋆ This work under
is NRI grantin1924721.
supported part by the U.S. National Science Fig.
Fig. 1.
1. (a)
trajectory
(a) Cable
Cable failure
tracking
failure to(b)
(b) Reconfigurability
continue
Reconfigurability allows
allows the
the
Foundation
This workunder NRI grantin1924721.
is supported part by the U.S. National Science
⋆
Foundation
This work under
is NRI grant
supported in 1924721.
part by the U.S. National Science Fig. 1. (a)
trajectory
trajectory Cable failure
tracking
tracking to
to(b) Reconfigurability
continue
continue allows the
Foundation under NRI grant 1924721.
2405-8963 Copyright © 2022 The Authors. This is an open access article under the CC BY-NC-ND license.
Foundation under NRI grant 1924721. trajectory tracking to continue
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2022.11.211
370 Adhiti Raman et al. / IFAC PapersOnLine 55-37 (2022) 369–375
driven parallel manipulator that controls a payload over
a large workspace (called the macro-manipulator); and
(ii) a continuum robot enclosing a concrete delivery tube
that provides precisely directed control through congested
spaces (called the micro-manipulator) (Srivastava et al.
(2022)). The presented work contributes to the develop-
ment of the first subsystem, for which we have chosen to
utilize a kinematically redundant CDPR. We presented
the modeling and development of kinematically redundant
CDPRs in our previous work (Raman et al. (2021)).
In this paper, we employ geometric redundancies with
redundancy resolution control to traverse a trajectory in
spite of cable failures. The proposed framework consists
of two parts: (i) a failure detection and identification
(FDI) section that introduces the use of an Interactive
Multiple Model Adaptive Filter; and (ii) a fault tolerant
control (FTC) section that utilizes a switching feedforward
kinematic controller in tandem with the estimator to form
a robust and automatic task recovery scheme.
2. RELATED WORK
The use of kinematic redundancy to allow serial rigid-link
robots to compensate for joint failures has had a long
history (Visinsky et al. (1994); Maciejewski (1990)). Early
studies accounting for cable failures in CDPRs (Roberts
et al. (1998)) showed that a static equilibrium could be
maintained (upon failure) if the end-effector has been
in specific singular configurations. Bosscher and Ebert-
Uphoff (2004) noted that there are two kinds of failure
modes: cable breakage due to excessive positive cable
tensions or slackness due to the lack of any tension. Notash
(2012) looked into additional cable failures, such as stuck
actuators (applying a passive restraint on the end-effector)
or situations in which the actuator moves but the output
is biased. However, these works are limited to kinematic
models and CDPRs with ideal and inelastic cables. Other
approaches (Passarini et al. (2019) and Boumann and
Bruckmann (2022)) presented interesting emergency stop
strategies (upon cable failure) in cable suspended robots
with the focus being on minimizing the safety risk instead
of the continued functioning of the robot.
Although FDI, FTC, or task recovery is a well explored
domain in aerospace applications, literature on the same
topics in the field of cable driven manipulators is almost
non-existent. A fairly popular general technique employed
for FDI in other domains is the use of Multi Model Adap-
tive Estimation (MMAE) filters that assume knowledge of
all the possible failure modes that may occur. In this study,
we use an Interactive MMAE or IMM for short. IMMs have
a rich history in trajectory tracking (Mazor et al. (1998))
and behaviour prediction (Gill et al. (2019)) as well as in
fault detection and identification (Tudoroiu and Khorasani
(2005)). Many of these studies focus on FDI only and
do not attempt to integrate controllers for task recovery.
One of the few exceptions is the work of Hill et al. (2021)
where a nonlinear Model Predictive Controller is applied
to recover from reaction wheel failures that are identified
through an UKF based IMM for satellite maneuvering.
Here, we employ a similar framework for FTC.
For our approach, the cables are modeled as elastic springs.
Furthermore, we assume that the employed dynamic mod-
Fig. 2. The 3 kinds of failures in a 4 cable kinematically
redundant CDPR
els are erroneous and do not correspond perfectly to phys-
ical reality (e.g., through parametric uncertainties and
process noise). We also assume that encoder readings do
not correspond accurately to the true position of the end-
effector and constitute an unreliable source of information;
a reasonable assumption to make when accounting for elas-
ticity in cables. There are several studies addressing the
ineffectiveness of relying solely on forward kinematics from
cable length measurements to determine end-effector pose,
such as works by Le Nguyen and Caverly (2021); Korayem
et al. (2018); Caverly and Forbes (2016) where techniques
based on Extended (EKF) and Unscented (UKF) Kalman
Filtering based on data from payload mounted IMUs have
been implemented to mitigate this issue.
The IMM applies a parallel bank of filters, each corre-
sponding to one of the various failure modes. The out-
put provides an estimate of model corrected pose to-
gether with an understanding of the current working mode.
Then, the controller computes a mixed joint input to the
plant/environment based on real-time information from
the IMM. This approach simultaneously detects the fault,
provides a corrected estimate, and updates the input, such
that the system can robustly cope with sudden failures.
This paper serves as a proof-of-concept with key contri-
butions as follows: (i) an Interactive Multi Model Filter
is derived and demonstrated as a robust estimator for
cable failure detection and identification (FDI) as well
as for real-time pose estimation in CDPR applications;
and (ii) a first iteration of a proportional-derivative based
fault-tolerant control algorithm for task recovery through
redundancy resolution post FDI is demonstrated.
3. FAILURE MODES
The loss of cables in a CDPR reduces the degrees of
freedom (DOFs) or the quality of the DOF available at the
end-effector. We consider 2 DOFs at the end-effector for
the four-cable planar CDPR model in this paper. The total
working modes considered can be described by 3 motion
models: (i) an over-constrained CDPR with 4 cables; (ii) a
fully constrained CDPR with 3 cables; and (iii) an under-
constrained CDPR with two cables (Fig 2).
Therefore there is a total of 7 working modes: no cables
have failed (mode 1), only cable A, B, C, or D fail (modes
2-5), cables A and B have failed (mode 6), or cables C and
D have failed (mode 7). If any other failure combination
occurs, the robot becomes non-navigable for significant
areas of the workspace.
 Adhiti Raman et al. / IFAC PapersOnLine 55-37 (2022) 369–375
Fig. 3. The kinematically redundant 4 cable CDPR
4. CDPR DYNAMICS
This section discusses the dynamic model of a planar
kinematically redundant CDPR with a 2-DOF platform
that is driven by 4 cables and 4 linear actuators housing
the cable winch mechanism. The full kinematic model
has been presented in our previous work (Raman et al.
(2021)). Typical CDPRs require a greater number of cables
than the degrees of freedom at the end-effector to be
fully constrained. In our model, the cable elements are
considered as linear springs with the specific stiffness
(stiffness per unit free cable length) of the cable given by
k0,i . Each cable link forms a prismatic joint between the
end-effector {F }i and the base {S3 }i attachment points.
The joint length is then given by lp,i which is actuated by
a motor-encoder whose position is given by θi . This term
should ideally be rθi where r is the motor winch radius,
but since r is a constant, we subsume the entire term under
a generic notation θi . Due to the elasticity, lp,i ̸∝ θi . The
cable stiffness at the current time step is given by ki =
k0,i /lp,i , where lp,i is determined from inverse kinematics
under the assumption of an available end-effector pose
(for the controller this comes from forward kinematics).
The tension in the cables is given by ki (lp,i − θi ), where
lp,i > θi to prevent cable sag. The geometric attachment
points, {S2 }i , are not constant as in traditional CDPRs,
but instead move along linear actuators along the x-axis of
the base frames {S1 }i . The longitudinal positions of these
attachment points with respect to the base frame are given
by ls,i . For our model, the linear sliders are considered
to be perfect without providing any additional dynamic
effects (subject only to limits on velocity). The complete
joint input vector can then be expressed as q = [ls θ].
Note that the lower level control signals translating joint
inputs to actuator torque occur independent of the CDPR
control loop and are therefore not modeled here.
The cable failure is modeled as a sudden drop in the cable
stiffness of the failed joint, i.e., ki → 0 over a small time
interval of ∆t = 0.1s. The full dynamic model of the
system can then be derived via the Lagrangian approach
as
M ẍe + Dẋe = P (xe , ls ) Kq lp (xe , ls ) − θ (1)
where xe = [x, y, ϕ] is the end-effector pose. P (xe , ls ) is the
pulling map or wrench Jacobian. The prismatic length lp
comes from inverse kinematics. Both are dependent on the
current pose xe and geometric attachment points ls while
371
Fig. 4. The Interactive Multiple Model (IMM) filter
the joint state stiffness matrix Kq is a diagonal matrix with
the elements (k1 , . . . , k4 ). The mass and damping matrices
are M and D respectively. This formulation ignores the
effects of gravity (as the system is planar).
5. MOTION MODELS
Three motion models are employed to describe the seven
working modes. For each working mode, the effect of the
corresponding cable failure is incorporated in the dynamic
model. The states of the end-effector are given by x =
T
[xe ẋe ] . The resulting non-linear system model is given
by
ẋ(t) = f(x(t), u(t), t) + ω
ỹ = h(x(t), t) + v (2)
where ω and v are zero mean Gaussian vector processes
accounting for model and measurement noise. The input
u(t) to the system is given by the joint input position
vector q(t). The non-linear vector functions f and h emerge
as
 
 ẋe 
f(x, u) =
M −1 P (xe , ls )Kq lp (xe , ls ) − θ − M −1 Dẋe
 
x
h(x) = e (3)
0
where the explicit dependence on time has been dropped
for ease of notation. The measurement model assumes that
noisy position data is directly available, for instance by
separately pre-processing and filtering information from
an on-board IMU.
6. INTERACTIVE MULTIPLE MODEL ESTIMATION
An interactive multiple model filter (IMM) is a dynamic
estimator (Blom and Bar-Shalom (1988)) which can be
used when model changes appear suddenly or gradually
over time, thus providing means for failure detection. The
process flow of an IMM is shown in Fig. 4. In general, the
bank of filters in an IMM can be realized by a variety of es-
timation approaches, i.e., Kalman Filters (KF), Extended
Kalman Filters (EKF), Unscented Kalman Filters (UKF),
Particle Filters (PF), et cetera. In this initial study, we
employ an EKF formulation in which the error dynamics
are approximated by a first order Taylor series expansion
about the estimated state, x̂, i.e.
∂f(x, u)  ∂h(x) 
F =  , H=  (4)
∂x x̂ ∂x x̂
372 Adhiti Raman et al. / IFAC PapersOnLine 55-37 (2022) 369–375

Here, F and H can be determined numerically by a 1

complex step derivative approximation. In this study, the p(ỹk |x̂−j
k )=
[det (2πEk−j )]1/2
EKF is applied in the discrete time domain; hence, the  1 
Euler-Maruyama discretization of the motion model yields · exp − e−jTk Ek−j −1 e−j
k
√ 2
xk = xk−1 + f(xk−1 , qk−1 )∆t +ω k ∆t with Ek−j = Hkj Pk−j HkjT + Rk
  
ϕ(xk−1 , qk−1 ) e−j −j
k = ỹk − ŷk (10)
ỹk = h(xk ) + vk (5) and where e−j
k is the estimation error.
(3) Model Probability Update: Now, the model probabil-
where ỹ ∈ R is the measurement vector, ω k ∼ N (0, Qk )
m
ities are updated via the likelihood with subsequent
denotes the process noise with covariance Qk ∈ Rn×n , and
normalization, i.e.
vk ∼ N (0, Rk ) is the measurement noise with covariance
Rk ∈ Rm×m . The sampling time is denoted by ∆t, and k wkj = wk−1
j
p(ỹk |x̂−j
k )
is the current time step. wj
wkj ← Mk (11)
Let M be the number of all working modes in a CDPR wkj i=1
at the k th time step. Then, the input to the filter is the
(4) Combination: Finally, the updated estimate is given
current measurement and mixed state estimate x̂+0 . The
by
algorithm broadly consists of four stages (Gill (2019)):
M

(1) Interaction and Mixing: The weights and the esti- x̂+ wkj x̂+j
k = k
mates from the last cycle are mixed as per their asso- j=1
ciated Markov transition probabilities. The predicted M   T 
probability for the filter to end up in mode j in each 
Pk+ = wkj x̂+j +
k − x̂k x̂+j +
k − x̂k + Pk+j
cycle, given that it was in the mode i during the
j=1
previous cycle is given by
(12)
(i|j) 1 (i)
wk = w pij
j k The final estimate of x̂ is not integrated into the overall
c̄k controller but the weight vector, w, plays an important
M
 (i) role. The weight vector wkj displays the importance (proba-
c̄jk = wk pij (6) bility) associated with each model in the bank. If there are
i=1 no cables failures, for example, the IMM will determine the
with c̄jk being the normalization factor, and where pij first model (mode 1) to have the largest weight. This vector
is the Markov transition probability from mode i to also informs the balance of joint inputs to the controller
mode j. This matrix is tuning factor and is deter- as described in the next section.
mined heuristically. The mixed initial state estimate
at the start of the current time step is provided by 7. TASK RECOVERY
M

x̂+0j = wk
(i|j) +i
x̂k We wish to describe a proof-of-concept controller for a
k
i=1
CDPR performing 3D printing. Therefore, the primary
objective is to maintain end-effector trajectory tracking
M
  
(i|j) along a predefined path within the operational workspace
Pk+0j = wk x̂+i +0j
k − x̂k of the redundant CDPR despite cable failure. The oper-
i=1
 T  ational workspace is different from the static workspace.
x̂+i
k − x̂+0j
k + Pk+i (7) The static workspace is the instantaneous wrench feasible
workspace of the CDPR if all the sliders were fixed. The
(2) Model Specific Filtering: The mixed initial estimates operational workspace is the overall traversable workspace
are fed into the EKF and processed in two steps, i.e. of the CDPR that lies within the total geometric con-
• Propagation: straints of the system. For the reconfigurable CDPR, the
operational workspace lies within the bounds of the blue
x̂−j +0j
k = x̂k + f(x̂+0j
k , qk )∆t box in Fig. 1 (geometric limits of the joints), but we cannot
make declarations about the quality and existence of the
Pk−j = Φjk Pk+0j ΦjT
k + Qk static workspace (for all redundant configurations) within
Φj = I + Fkj ∆t (8) it with equal ease. However, it is sufficient to state that if
• And update: the next pose in the trajectory is manipulable and wrench
feasible, then it is realizable through the controller.
(j)
Kk = Pk−j HkjT [Hkj Pk−j HkjT + Rk ]−1 The desired trajectory of the end-effector is known. The
−j −j
x̂+j
k = x̂k + Kk [ỹk − h(x̂k )]
corresponding joint velocities are determined through re-
dundancy resolution. The primary motivation of the task
Pk+j = [I − Kk Hk ]Pk−j (9)
recovery controller is to reduce the tracking error as a pri-
Here, ‘+’ denotes the a posteriori estimate whereas mary constraint while maintaining the manipulability and
‘−’ denotes the a priori quantity before the update. wrench feasibility (i.e., avoiding infeasibility and singular-
The likelihood of a measurement is then given by ity) of the joint space solutions. In general, this process
 Adhiti Raman et al. / IFAC PapersOnLine 55-37 (2022) 369–375 373

can be separated into: (i) a task space controller; (ii) a

redundancy resolution scheme to maximize manipulabil-
ity and wrench feasibility; and (iii) a joint level velocity
controller.

7.1 Trajectory Tracking

The first iteration for the task recovery task space con-
troller was a straightforward discrete-time Proportional-
Derivative (PD) based trajectory tracking algorithm, i.e.
ekp = (xke,r − xke,f )
ekd = (ekp − ek−1
p )/∆t
xk+1
e = xke,f + gp ekp + gd ekd
where xke,r is the trajectory point at current time step
and xke,f comes from forward kinematics. The forward
kinematics are determined from a Levenberg-Marquardt
minimization of potential energy in the system given the
joint states at the current time step. The next desired
position for the end-effector, xk+1 , is employed to calculate
e
the required joint states. The control gains gp and gd are
selected as [0.6 0.6 0] and [0.1 0.1 0], respectively. We do
not control motion in the third DOF.
7.2 Redundancy Resolution
For a given end-effector pose, there is an infinite amount of
solutions in the joint space for a kinematically redundant
CDPR. The optimal joint slider positions are determined
from an objective function that seeks to maximize the
manipulability ellipsoid at the end-effector. This aids in
avoiding singularities as the sliders travel. When a cable
fails (second motion model), the manipulability in the
third DOF is ignored. The joint sliders are velocity limited
to vmax and vmin in either direction in order to avoid
discontinuous motions i.e.,
l∗s = max κ
ls
Fig. 5. Automatic task recovery through open-loop kine-
matic control
(13) 7.3 Feedforward Kinematic Control
The task recovery algorithm has a parallel bank of trajec-
tory tracking controllers, each corresponding to a working
mode. The controllers use simple kinematic models for the
different redundant CDPR motion modes. These inputs
are mixed and normalized with the weight vector arising
from the IMM, giving precedence to the input correspond-
ing to the detected failure mode. The bank of task- and
joint-space controllers each accept the mixed joint state
input of the previous time-step and propagate it forward
to the next input for each failure mode. The mixed joint
inputs are then applied to the dynamic simulation model
(plant), and the resulting (noisy) measurements of the end-
effector state inform again the state estimation and weight
vector of the IMM.
The dynamic simulation of the plant and IMM state esti-
mation utilize a sampling rate different from the controller.
The mixing process results in a gradual change of the joint
inputs to the plant, thus allowing for recovery if a failure
mode is incorrectly detected. The complete algorithmic
framework is shown in Fig. 5.
8. IMPLEMENTATION AND RESULTS

lk+1
s = k
ls + ∆ls 8.1 Simulation Framework
∆ls = min(max(l∗s
− lks , vmax ∆t), vmin ∆t)
(14)
The plant is a dynamic model created in MATLAB which
where κ is a measure of the sensitivity of the pulling
is perturbed with process noise to mimic unmodeled dy-
map at the desired pose. This is an equivalent property
namics. The measurements are simulated though a mea-
to the measure of manipulability. This value is given by
min z surement noise perturbation as well. The cables are con-
κ = max z , where z = null(P ). The closer κ is to 1, sidered as linear springs (where stiffness is a function of
the better conditioned the pulling map is. For the final
free cable length) with no mass and no sag while all other
motion model, the DOF at the end-effector reduces to one
model parameters are constants. Cable failure is modeled
and the manipulability ellipse reduces to a line making
as a drop in cable stiffness to zero. If the stiffness is zero,
singularity avoidance a non-issue. For this case, since
the cable will have no tension and thus have no impact on
the cables can control only once DOF, the sliders utilize
the platform dynamics. The plant and estimation run at
the same controller for trajectory error minimization to
100 Hz while the kinematic controller runs at 10 Hz.
determine actuation that will manipulate the end effector
in the remaining DOF.
8.2 Task Recovery
The joint angles θ are determined by first extracting the
desired tensions in the cables, i.e., given the desired end- Figure 6 illustrates the application of the automatic task
effector position together with lk+1s , we determine the recovery algorithm for a case when a first cable breaks
minimum positive tensions required to maintain this pose: at 5 seconds and an additional cable fails at 10 seconds.
τ ∗ = min τ T τ ∋ P (xk+1 , lk+1 )τ = 0 (15) The red trajectory shows the true path of the end-effector
e s
τ while the blue trajectory depicts the expected path. The
The joint angles are then determined in a straightforward red line exhibits a clear break away from the trajectory
fashion by and subsequent recovery with three cables as the systems
θ = lk+1
p − Kq−1 τ ∗ (16) returns to and continues from the point of first failure. The
374 Adhiti Raman et al. / IFAC PapersOnLine 55-37 (2022) 369–375
Fig. 6. Trajectory tracking and recovery
Fig. 7. Cable failure identified at 5 and 10 seconds
Fig. 8. Slider positions during geometric reconfiguration
Fig. 9. Trajectory error correction during task recovery
second cable failure and subsequent recovery then demon-
strates similar behavior further down the trajectory. When
the cables break, the IMM responds instantly, identifying
the correct model and mixing the inputs accordingly, thus
leading to recovery. In Fig. 7, the weights vector correctly
assigns the largest values to the current working mode. At
the five second mark, we can see that the weight vector
has correctly identified the cable A failure, followed by
the correct identification of the cable C failure at the
ten second mark. The slider positions change in time to
reflect the redundancy resolution during task recovery in
Fig. 8 while the robustness of the task recovery can be
seen in Fig. 9 as the real and desired trajectories converge.
Working in simulation provides us with access to the
ground truth and thus, the estimation error (e = xe − x̂e ),
therefore allowing for insight into the stability of the IMM.
Fig. 10. Estimation Error
8.3 Drawbacks
When the desired trajectory lies outside of the new static
workspace that forms just after cable failure, the robot
sometimes struggles to recover the trajectory tracking.
This is because the robot may move close to singularity
while attempting to minimize the trajectory error too
quickly. Sometimes cable failure can also lead to cable
slackness. While this property is not explicitly modeled in
the system dynamics, the destabilizing effect of this can be
identified when the value of spool angles, θ, is too high (at
the instant of failure) to converge to a forward kinematics
solution that lies within the new static workspace. For 3D
printing tasks, we also require the end-effector to return
to the point of break. These drawbacks will be addressed
in our next work.
9. CONCLUSION
In this study, we demonstrate for the first time a proof-
of-concept approach to failure detection and identifica-
tion (FDI) with fault tolerant control (FTC) for kine-
matically redundant CDPRs by automating the cycle of
detection and recovery. CDPRs offer a unique advan-
tage by being able to utilize geometric reconfiguration
in overcoming the consequences of cable failure. Failure
modes are detected/estimated via an Interactive Multiple
Model (IMM) algorithm that relies on a bank of parallel
Extended Kalman Filters. The actual joint state inputs
are accomplished by mixing the simultaneously calculated
inputs from a parallel bank of pre-developed trajectory
tracking controllers utilizing the IMM weight vector. While
IMM approaches have been employed in dynamic fault
detection and identification for aerospace applications, its
use for fault detection in the robotics domain in general is
sparse and truly unique for the cable driven robotics com-
munity. Furthermore, the recovery strategy incorporating
both redundancy resolution and geometric reconfiguration
for task recovery with the utilization of the IMM weights
for control mixing is a first in this domain. In addition
to providing a novel research idea, this also serves as a
vehicle demonstrating the powerful advantage of adding
kinematic redundancies to CDPRs. This work is meant to
be a first study of a concept that requires advancement and
 Adhiti Raman et al. / IFAC PapersOnLine 55-37 (2022) 369–375
refinement in future work, for instance by incorporating
a closed loop dynamic controller together with a motion
planning strategy to recover the trajectory immediately
from the point of failure. Further exploration is suggested
on how to incorporate redundant actuation to minimize
dynamic effects of cable failure to avoid cable and object
collision.
ACKNOWLEDGEMENTS
This work is supported in part by the U.S. National
Science Foundation under NRI grant 1924721.
REFERENCES
Blom, H.A. and Bar-Shalom, Y. (1988). The interacting
multiple model algorithm for systems with markovian
switching coefficients. IEEE transactions on Automatic
Control, 33(8), 780–783.
Bosscher, P. and Ebert-Uphoff, I. (2004). Wrench-based
analysis of cable-driven robots. In IEEE International
Conference on Robotics and Automation, 2004. Proceed-
ings. ICRA’04. 2004, volume 5, 4950–4955. IEEE.
Boumann, R. and Bruckmann, T. (2022). Simulation and
model-based verification of an emergency strategy for
cable failure in cable robots. In Actuators, volume 11,
56. MDPI.
Caverly, R.J. and Forbes, J.R. (2016). State estimator
design for a single degree of freedom cable-actuated
system. Journal of the Franklin Institute, 353(18), 4845–
4869.
Chesser, P.C., Wang, P.L., Vaughan, J.E., Lind, R.F., and
Post, B.K. (2022). Kinematics of a cable-driven robotic
platform for large-scale additive manufacturing. Journal
of Mechanisms and Robotics, 14(2).
Gagliardini, L., Caro, S., Gouttefarde, M., Wenger, P., and
Girin, A. (2015). A reconfigurable cable-driven parallel
robot for sandblasting and painting of large structures.
In Cable-Driven Parallel Robots, 275–291. Springer.
Gill, J.S., Pisu, P., Krovi, V.N., and Schmid, M.J. (2019).
Behavior identification and prediction for a probabilistic
risk framework. In Dynamic Systems and Control
Conference, volume 59155, V002T25A004. American
Society of Mechanical Engineers.
Gill, J.S. (2019). Probabilistic Framework for Behavior
Characterization of Traffic Participants Enabling Long
Term Prediction. Ph.D. thesis, Clemson University.
Hill, E., Gadsden, S.A., and Biglarbegian, M. (2021). Ex-
plicit nonlinear mpc for fault tolerance using interacting
multiple models. IEEE Transactions on Aerospace and
Electronic Systems, 57(5), 2784–2794.
Izard, J.B., Dubor, A., Hervé, P.E., Cabay, E., Culla, D.,
Rodriguez, M., and Barrado, M. (2017). Large-scale 3d
printing with cable-driven parallel robots. Construction
Robotics, 1(1), 69–76.
Izard, J.B., Gouttefarde, M., Baradat, C., Culla, D., and
Sallé, D. (2013). Integration of a parallel cable-driven
robot on an existing building façade. In Cable-driven
parallel robots, 149–164. Springer.
Jamshidifar, H., Fidan, B., Gungor, G., and Khajepour,
A. (2015). Adaptive vibration control of a flexible cable
driven parallel robot. IFAC-PapersOnLine, 48(3), 1302–
1307.
375
Korayem, M., Yousefzadeh, M., and Kian, S. (2018).
Precise end-effector pose estimation in spatial cable-
driven parallel robots with elastic cables using a data
fusion method. Measurement, 130, 177–190.
Le Nguyen, V. and Caverly, R.J. (2021). Cable-driven
parallel robot pose estimation using extended kalman
filtering with inertial payload measurements. IEEE
Robotics and Automation Letters, 6(2), 3615–3622.
Maciejewski, A.A. (1990). Fault tolerant properties of
kinematically redundant manipulators. In IEEE Inter-
national Conference on Robotics and Automation, 638–
642.
Mazor, E., Averbuch, A., Bar-Shalom, Y., and Dayan, J.
(1998). Interacting multiple model methods in target
tracking: a survey. IEEE Transactions on aerospace and
electronic systems, 34(1), 103–123.
Notash, L. (2012). Failure recovery for wrench capability
of wire-actuated parallel manipulators. Robotica, 30(6),
941–950.
Passarini, C., Zanotto, D., and Boschetti, G. (2019). Dy-
namic trajectory planning for failure recovery in cable-
suspended camera systems. Journal of Mechanisms and
Robotics, 11(2).
Raman, A., Schmid, M., and Krovi, V. (2020). Stiffness
modulation for a planar mobile cable-driven parallel
manipulators via structural reconfiguration. In Inter-
national Design Engineering Technical Conferences and
Computers and Information in Engineering Conference,
volume 83990, V010T10A054. American Society of Me-
chanical Engineers.
Raman, A., Schmid, M., and Krovi, V.N. (2021). Wrench
analysis of kinematically redundant planar cdprs. In In-
ternational Conference on Cable-Driven Parallel Robots,
90–104. Springer.
Rasheed, T. (2019). Collaborative Mobile Cable-Driven
Parallel Robots. Ph.D. thesis, École centrale de Nantes.
Roberts, R.G., Graham, T., and Lippitt, T. (1998). On the
inverse kinematics, statics, and fault tolerance of cable-
suspended robots. Journal of Robotic Systems, 15(10),
581–597.
Seriani, S., Gallina, P., and Wedler, A. (2016). A modular
cable robot for inspection and light manipulation on
celestial bodies. Acta Astronautica, 123, 145–153.
Srivastava, M., Ammons, J., Peerzada, A., Krovi, V.,
Rangaraju, P., and Walker, I.D. (2022). 3d printing
of concrete with a continuum robot hose using variable
curvature kinematics. In IEEE International Conference
on Robotics and Automation (ICRA). IEEE.
Tudoroiu, N. and Khorasani, K. (2005). Fault detection
and diagnosis for satellite’s attitude control system (acs)
using an interactive multiple model (imm) approach.
In Proceedings of 2005 IEEE Conference on Control
Applications, 2005. CCA 2005., 1287–1292. IEEE.
Visinsky, M.L., Cavallaro, J.R., and Walker, I.D. (1994).
Robotic fault detection and fault tolerance: A survey.
Reliability Engineering and System Safety, 46(2), 139–
158.
Zhou, X., Jun, S.k., and Krovi, V. (2014). Stiffness mod-
ulation exploiting configuration redundancy in mobile
cable robots. In 2014 IEEE International Conference on
Robotics and Automation (ICRA), 5934–5939. IEEE.