RRR-ROBOT: VALIDATING NONLINEAR CONTROL CONCEPTS IN

AN INDUSTRIAL-LIKE, 3 DOF, MANIPULATOR

Bert van Beek
Faculty of Mechanical Engineering, Eindhoven University of Technology,
Eindhoven, The Netherlands, bert@wfw.wtb.tue.nl
Bram de Jager
Faculty of Mechanical Engineering, Eindhoven University of Technology,
Eindhoven, The Netherlands, A.G.de.Jager@wfw.wtb.tue.nl

ABSTRACT
A new experimental facility, the RRR-robot has been designed and build to compare and validate a variety of
advanced nonlinear control strategies. To achieve a resemblance both in appearance and in complexity with
conventional industrial robots (e.g., Puma-type robots),
the system has a chain structure and three rotational degrees of freedom. The design concept and initial experimental evaluation of the RRR-robot are addressed. Preliminary experimental results with (model) based control
laws will be presented.
Keywords: robot design, direct-drive motor, nonlinear
control system.
1 INTRODUCTION
Industrial robots. Robots are applied in a variety of
fields, of which industrial automated manufacturing is an
important one. Objectives such as reducing the manufacturing costs, increasing the productivity, and improving (or maintaining) the product quality standards, represent the main factors that have promoted an increasing
use of robotics technology. Nowadays, typical robot applications include: materials handling (e.g., palletizing
or packaging), manipulation (e.g., arc welding or spray
painting), and measurement (e.g., object inspection).
There is an increasing demand from industry for systems which can achieve these tasks faster and/or more accurately. If the increase of performance can be achieved
at the same price, both improvements will result in lower
price per product since increased accuracy may result in
a reduction of post processing steps, and the increase of
velocity increases throughput.

Contacting author: P.O. Box 513, 5600 MB Eindhoven, Tel:

+31 40 2472784, Fax: +31 40 2461418.

Performance limitations. However, several unwanted

effects may limit or reduce the performance: configuration dependent inertia and conservative terms, Coriolis and centrifugal torques, friction, flexibilities in joints
and/or links, backlash, motor dynamics, and sensor dynamics. Various control schemes have been developed to
deal with one or more of these effects.
For high-speed tracking of complex trajectories, the Coriolis and centrifugal torques form an essential part of the
occurring nonlinear effects. Therefore, these terms play
an important role in many of the proposed control strategies [4, 5]. For further development it is necessary to
test and compare these concepts in an experimental setup
with a complexity comparable to industrial manipulators
with multiple degrees of freedom [1].
Objective RRR-robot. The main problem is to enhance
the Coriolis and centrifugal torques relative to other linear
and nonlinear effects. In an industrial manipulator these
effects may obscure the influence of the velocity dependent torques. The objective of the RRR-robot is to solve
this problem, i.e., to highlight the influence of Coriolis
and centrifugal torques in an experimental facility with
an industrial-like complexity.
Outline of this paper. In Section 2, the problem which
the RRR-robot aims to solve is analyzed, and all relevant
design specifications are summarized. Next, in Section 3,
the robot is conceptually divided into four subsystem: a
Joint-Actuation System, a Measurement System, a Control System, and a Signal-Transfer System. Each of these
subsystems is discussed in detail, and working solutions,
or even components are chosen. Subsequently, the subsystems are combined to form a complete system. In Section 4, an initial experimental evaluation is performed.
Finally, in Section 5, the results are summarized, and an
outlook of future work is presented.

2 DESIGN SPECIFICATIONS
2.1 Highlighting Coriolis and centrifugal torques
In order to highlight the Coriolis and centrifugal torques
their contribution to the total torque should be substantial during a sufficiently long period of time. To quantify this requirement, a performance measure is defined,
R
| {C q}
i | d
ri = R
,
(1)
| {M q}
i + {C q}
i + {g}i + {v}i | d
based on the familiar rigid robot model in the joint space:
M(q) q + C(q, q)
q + g (q) + v (q)
=

(2)

where q denotes the vector with rotation angles, q the angular velocities, and q the angular accelerations. So the
actuator torques consist of inertia terms, M(q) q,
Coriolis and centrifugal torques, C(q, q)
q,
torques derived
from potential fields like gravitation, g (q), and friction
torques, v (q).

Maintaining the Coriolis and centrifugal torques at a

significant level during a considerable period corresponds
to achieving a large ratio ri .

where K r is a diagonal matrix containing the gear reduction ratios kri . The central moments of inertia, Im i ,
of each motor with respect to its rotor axis are collected
in the diagonal matrix Im . Then, the mass matrix M(q)
can be split into a constant diagonal matrix, K r Im K r and
a configuration-dependent (dependent on functions of q)
matrix 1M(q), i.e.,
M(q) = K r Im K r + 1M(q).

Combining (2), (3) and (4) yields the block scheme of

Fig. 1. As can be seen the system described by Fig. 1
actually consists of two subsystems:
a linear and decoupled subsystem in which each component of m influences only the corresponding component of q m , and
a nonlinear and coupled subsystem.
Nonlinear and coupled
K r1 (g + v)

2.2 Problem analysis

In most industrial manipulators, the value of ri is limited
by rotational constraints and the presence of gears with
high transmission reduction ratios. In the following Sections both limitations will be discussed.
Constrained rotations. As can be seen in (1), a combination of high velocities and especially low accelerations
is a prerequisite for a large ratio ri . If the joint rotation angles are limited (e.g., by cables used for energy and data
transport), such a combination can be achieved only with
considerable energy (using powerful, expensive servos)
and only over a short period of time, since shortly after
the acceleration to a certain velocity, the de-acceleration
must begin in order to avoid breaking any cables. Furthermore, the large injection of energy will excite additional
vibrational modes, thus clouding the contribution of the
Coriolis and centrifugal torques.
Transmissions. The second obstacle in achieving a large
ratio ri is the presence of gears with high reduction ratios (e.g. harmonic drives). As will be explained, this has
the effect of linearizing and decoupling the system, and
thus reducing the ri . Next to this for conventional robot
desirable effect, gear boxes are a source of additional
friction, backlash and elasticity.
To illustrate their linearizing effect, the transmissions
are included in the model. Assume, the transmissions
are rigid and without backlash. Let q m and m denote
the vectors of actuator displacements and actuator driving torques respectively; the transmissions then establish
the following relationships:

q m = Kr q
(3)
m = K r1

(4)

K r1 C K r1

K r1 1M K r1

+
m

Im1

R
q m

R
q m

qm

Linear and decoupled

FIGURE 1: Block scheme of a rigid manipulator
with transmissions.

In case of high reduction gears (kri  1), the contribution

of the nonlinear interaction term d is small, and can be
considered as a disturbance for each joint servo. As a
result the ratio of (1) becomes small, since the torque m i
is almost completely determined by the motor inertia.
By choosing servos without transmissions (K r = I ),
direct-drives, the influence of nonlinearities and coupling
between the joints becomes significant, while eliminating
gear boxes as a source of additional friction, backlash and
elasticity.
2.3 Design requirements
The main requirement of the RRR-robot is the ability to
maintain the Coriolis and centrifugal torques sufficiently
large. It was decided that at least one servo should have
an ratio ri 0.5. From the problem analysis follows that
this can be achieved by the removal of constraints on the
rotation of each joint, and the use of direct-drive servos.

Other design requirements are:

a direct control of the actuator torque;
a modular set-up: the third link and joint should be exchangeable by, or expandable to flexible components;
three or more rotational DOFs (to achieve a resemblance with industrial-like manipulators);
an end-effector resolution of 0.1 [mm], allowing for
a gradual degradation to 1 [mm] for reference inputs
from 5 [Hz] up to 25 [Hz]; and
a control interface integrated with Matlab/Simulink.
3 RRR-ROBOT SUBSYSTEMS
To tackle the design problem, a choice is made to conceptually divide the RRR-robot into several subsystem
based on their function (see Fig. 2). To a certain extent,
these subsystems can be considered independently. In the
following, all subsystems, and their interactions are defined.
Measurement System
power signals

joint
sensors

measurement/control signals

power
supplies

power
amplifiers

servo
motors

transmissions

link
sensors

cartesian
position

robot
joints
and
links

Joint Actuation System

Signal
Transfer
System
Control
System

FIGURE 2: A possible functional lay-out of the RRRrobots subsystems, and their interaction. Note that, especially the position of the Signal-Transfer System relative
to the Joint-Actuation System may vary (e.g., depending
on the type and placement of the power amplifier).

The Joint-Actuation System consists of servo motors

with appropriate power supplies and power amplifiers, but without transmissions.
The Measurement System consists of
joint sensors (e.g., to measure joint-angles),
link sensors (e.g., strain gauges), and
a 3D measurement system to determine the Cartesian position of the end-effector independently
from the joint sensors.
The Control System consists of
hardware components for data acquisition, command output, and overload protection;
software components to implement the control algorithm, and user interface; and
a hardware platform to execute the software.
The Signal-Transfer System consists of components
for the transport of power and measurement signals,
across parts which move relative to each other.

TABLE 1: Main system parameters used in evaluation model.

The links are assumed to be rectangular aluminum beams. The
motor data are taken from the Dynaserv series of Litton Precision Products.
parameters link/joint
1
2
3
link length
[m]
0.5
0.2
0.2
link mass
[kg]
9.86
1.21
0.73
motor mass
[kg]
12
7.5
5.5
motor inertia [kg m2 ] 0.023 0.015 0.012
max. torque
[Nm]
60
30
15

3.1 Joint-Actuation System

Direct-drive type. Several direct-drive types were considered. Hydraulic motors have an excellent power-tomass ratio but present problems in combination with the
required unconstrained rotational freedom; pollution of
the working environment due to oil leakage is difficult
to avoid. In addition, high frequency pressure dynamics
complicate control of the output torque.
The alternative is the class of electrical servos. Although asynchronous (induction) motors are becoming
increasingly popular, they are typically suited to deliver
high power. Moreover, torque control is still difficult.
Therefore, a choice was made to use synchronous (permanent magnet) servos also known as brushless DC
servos which combine the controllability of the conventional DC motor with some of the advantages of
asynchronous motors: no mechanical commutation (thus
eliminating additional noise, friction and wear) and easy
heat disposal.
In a brushless DC motor, the permanent magnet rotor
position must be detected or measured, and fed back to
a driver which determines the appropriate stator phases.
Preferably, these internal sensors should be fit for controlling the robot.
Servo sizing. Choosing appropriate servos for the RRRrobot is an iterative problem, since, as a result of the required torque, each motor has also a significant mass.
To accelerate these masses, even larger motors may be
necessary. Therefore, to aid the motor choice, a rigidmanipulator model of the RRR-robot is used. This dynamic model is derived with the Euler-Lagrange formulation, using the Denavit-Hartenberg matrix representation [7] to describe the spatial displacement between the
neighboring link coordinate frames. For optimal flexibility Mathematica [9] was used to generate the resulting
equations of motion. Based on an initial design sketch,
a rough estimate was made for the link parameters (i.e.,
dimensions and material) leading to the main system parameters shown in the upper part of Table 1. To illustrate this initial design the kinematic relations were used
to draw the simplified wire-frame picture of Fig. 3. After
choosing an appropriate motion pattern, i.e., with a high

ratio r , a commercial motor can be evaluated by substituting its mass and rotor inertia. With the Dynaserv series of
BDC motors of Litton Precision Products (see the motor
data in the lower part of Table 1), the main design requirement, maintaining high Coriolis and centrifugal forces,
has been met (in Fig. 3, their contribution is illustrated).
Neglecting motor friction, the evaluation of (1) for each
servo yields r = [1, 0.39, 1].
Torques to achieve constant velocity

0.5

Motor 1 [Nm]

y
0
-0.5
0.8
z
0.4
0
-0.5

Motor 2 [Nm]

0
0.5
t= 0.13
0.5

y
0
-0.5

z
0.4
0
-0.5
x

0
50
0

(q1d , q d , q d )

C(q d , q d ) q d

20
0

0.5

3.3 Control System

For the Control System, a PC based solution was selected, using one or more PC CPUs (central processing
units) for both the controller and the graphical user interface (see Fig. 5, right). To achieve integration with
Matlab/Simulink, Windows was chosen as the operating
system (OS).
PC CPU

20
0

Motor 3 [Nm]

0.8

50

observer (camera or sensor) to the tip of the end-effector

is lost during normal operation.
Although several technologies were considered (e.g.,
acoustic ranging, laser tracking, and inertial measurement systems, see [8]), only one technology can satisfy
all requirements: vision based ranging using multiple
line-array cameras (see Fig. 4). By combining multiple
viewpoints also the line-of-sight problem can be solved.

GUI

5
0
5
0

ISA

GUI

DSP
controller

controller

ISA

1
time [s]

FIGURE 3: Left: wire frame animation based on kinematic relations, right: required total torque and the Coriolis and Centrifugal torques for a motion pattern with constant velocity q d = [ t, 2/3 t, 3 t]T [rad].

1
0
0
1
0
1
0
1
lense
0
1
0
1
0
1

111
000
000lense
111

RRRrobot

I/O
board

encoder

RRRrobot

A/D

FIGURE 5: PC related controller configurations: on the

left PC hosted, and on the right PC based. Both offer
Simulink integration, i.e., the control algorithm is implemented using Simulink block schemes.

Compared to a PC hosted controller with an embedded DSP (Fig. 5, left), this solution is less reliable; unpredictable behavior can occur if the OS crashes. Since,
Windows is not a hard-real time OS, the user is responsible for checking whether the desired sample frequency
is realized. However, when high level programming
(in C/Fortran or even using a automated-code generating
tool) is required, PC based controllers can offer a performance at least equal to a PC hosted solution at a much
lower price [3]. Based on the performance and price, a
choice was made to use a PC based controller: the MultiQ I/O board from Quanser Consulting [2], and a Pentium Pro 200.

right X-array
Y-array
FIGURE 4: 3D measurement system using 3 line-array cameras: cylindrical lenses compress a 2D image of the environment, and project it on CCD line-arrays to extract X, Y and the
range Z (from X l X r ) of a Light Emitting Diode.
left X-array

encoder
A/D

3.2 Measurement System

To evaluate the tracking behavior of an intentional elastic end-effector link or joint, an independent 3D tracking
system is required. The design specifications of this system, especially the combination of high accuracy (0.11 [mm]), high update rate (1 [kHz]) and uninterrupted
tracking, severely restrict the possible options. Since the
RRR-robot is designed to enable unconstrained joint rotation, it is inevitable that the line-of-sight from one single

LED

D/A

D/A
I/O
board

1111
0000
0000lense
1111

PC CPU

3.4 Signal-Transfer System

The Signal-Transfer System transports power and measurement signals across parts which move relative to each
other. Various transfer technologies exist; some only
suited for data transport (e.g., RF or optical), other suited
for both power and measurements signals (e.g., mechanical, magnetic or capacitive). However, because of the

signal requirements imposed by the selected direct driveservos, both for power and data transfer a conventional
mechanical solution was chosen: sliprings.
3.5 Synthesis
Combining all subsystems, the manipulator frame was
designed. The design started from the selected Dynaserv
motors with internal bearings and suitable joint anglesensors. Although restricting the designers freedom, the
use of complete servo-units greatly simplified and speedup the design process. Only for the first link, an additional bearing was necessary to reduce the moment load
on the first servo. The components of the slipring based
Signal-Transfer System are selected in conjunction with
the frame design based on the available space and available commercial products.

Ts
Tl

1
0

Te

clock
interrupts

11 0
00
1

1
0

latency
real-time
performance

FIGURE 7: Parameters defining the real-time performance.

power slipring
around servo

signal
slipring
inside

4.2 Real-time performance

Although the used OS (Windows) does not guarantee
hard-real time performance, the actual performance can
be (close to) real-time, depending on the conditions (e.g.,
the used hardware or the control complexity). In Fig. 7,
the parameters determining the real-time performance are
defined. Using an internal or external timer, interrupts are
generated with a desired sampling period Ts . When an interrupt occurs, there is a latency Tl before the foreground
tasks are stopped, and the real-time task is executed, resulting in a varying effective sampling period Te .

servo,
bearing &
encoder

FIGURE 6: Final design and realization of the RRRrobot.

In Fig. 6, the resulting design and realization of the

RRR-robot are shown. For the second servo (the white
servo), the locations of the power and signal sliprings
are indicated. The 3D measurement system for the endeffector has not yet been realized.
4 EXPERIMENTAL EVALUATION
4.1 Bench mark controller
To put the Control System to the test, a computational involved Passivity-based Computed Torque Controller [6]
is implemented, i.e.,
u = M(q)q d + C(q,
q)q d + g(q) + K D e + K P e (5)
with e = q d q the tracking error, and K D and K P
positive definite feedback matrices. The formulas for M,
C, and g were generated by Mathematica based on the
rigid-robot model.

The variation in Te can be determined by monitoring an

additional test output (a DA channel switching between
two voltage at every sample) on a scope. For the bench
mark controller, it was verified that an effective sampling
frequency of 1 [kHz] can be achieved without missing interrupts. Furthermore, the variation of Te is smaller than
0.05 [ms]. So, for the implementation of this computational involved (model based) control algorithm, the realtime performance seems adequate.
4.3 Servo performance
In an initial evaluation of the servos, their performance
proved to be less ideal than advertised:
The relation between command voltage and motor
output torque exhibits considerable saturation.
All motors have a Coulomb friction component of approximate 3 [Nm].
In Fig. 8 and Fig. 9 some initial experimental and simulation results with two controllers are displayed. All control parameters were tuned to achieve an acceptable, but
not necessarily optimal response. Without friction compensation, both controllers give considerable steady-state
errors. Furthermore, the friction reduces the contribution
of the Coriolis and centrifugal torque to the output torque,
especially for the lightest (end-effector) actuator. However, for the base servo, r1 > 0.5.
Comparing both controllers, it can be seen that the
feedforward M q d + C q d does not improve the tracking
behavior, probably due to incorrect model parameter values in the mass matrix M (and consequently C). A thorough experimental identification is required to adapt the
rigid-robot model parameters (now based on the robot design).

e2 [rad]

0.4
0.2

e1 [rad]

0.4

0.2

20

control inputs u(t)

u 1 [Nm]

measured
simulated

20

20

20

u 2 [Nm]

tracking errors e(t)

e2 [rad]

e1 [rad]

20

u 2 [Nm]

control inputs u(t)

u 1 [Nm]

tracking errors e(t)

0.4
0.2
0
0.2

20

0.4

0.2

0
6

time [s]

e3 [rad]

e3 [rad]

0.1

6
4
2
0
2

0.2

20

6
4
2
0
2

0.2

measured
simulated

0.1

time [s]

FIGURE 8: CTC-scheme: u = M q + C q + g +
d

K D e+
K P e, simulated and measured position errors
and control inputs without friction compensation.
The desired trajectory is a short (1 [s]), constant, acceleration to a constant velocity q d = [, 2, 3]T .
5 SUMMARY AND OUTLOOK
An industrial-like experimental facility, the RRR-robot,
has been realized that offers the ability to enhance Coriolis and centrifugal torques relative to other nonlinear effects. Two important features of the RRR-robot were introduced in order to reach this design goal: unconstrained
rotation of each link by using sliprings, and the use of
direct-drive servos. The robot has a modular setup, i.e., it
has been designed starting from a specific type of servo
with internal bearing, and angular sensors. The last link
can be replaced by a flexible link, and the last joint can
be transformed into a flexible joint. Thus making it suitable to evaluate control schemes for both rigid-robots and
flexible joint and/or link robots.
The system is equipped with a Simulink based graphical user interface, offering ease of use. The feasibility of
this Control System has been demonstrated by successfully implementing a model based control algorithm. Although the friction torques of the servos are considerable,
the contribution of the Coriolis and centrifugal torques
for at least one servo can be made sufficiently large.
After a thorough experimental identification of the
robot, and the implementation of the 3D measurement
system for the end-effector, future research may involve
the systematic evaluation of several control algorithms
for rigid, and flexible joint and/or link robots. Besides
encoder feedback, non-colocated actuation/sensing (using additional link sensors) and vision feedback schemes
can be implemented.
6 ACKNOWLEDGMENT
The project is financed by a grant of the board of Eindhoven University of Technology in the context of the participation in the Dutch Institute of Systems and Control.

time [s]

20

u 3 [Nm]

u 3 [Nm]

time [s]

FIGURE 9: PD controller with gravity compensation:

u = g + K D e + K P e (identical conditions and settings

as in Fig. 8).

REFERENCES
[1] W. J. Book. Modeling, design, and control of
flexible manipulator arms: A tutorial review. In
Proceedings of the 29th Conference on Decision and
Control, volume 3, pages 500506. IEEE, 1990.
[2] Quanser consulting. Wincon and MultiQ.
http://www.quanser.com, 1998.
[3] N. Costescu and D. Dawson. QMotor 2.0 - A PC
based real-time multitasking graphical control
environment. Submitted to IEEE Control Systems,
1997. Available at http://ece.clemson.edu/crb/
research/qmotor/qmotor2.pdf.
[4] Bram de Jager and Jos Banens. Experimental
evaluation of robot controllers. In Proc. of the 33rd
IEEE Conf. on Decision and Control, volume 1,
pages 363368, Lake Buena Vista, FL, 1994. IEEE,
Piscataway, NJ.
[5] I. M. M. Lammerts, F. E. Veldpaus, J. J. Kok, and
M. J. G. Van de Molegraft. Adaptive computed
torque computed reference control of flexible robots.
In ASME 1993 Winter annual meeting, New
Orleans, Louisiana, USA, December 1993.
[6] Ivonne M. M. Lammerts. Adaptive computed
reference computed torque control of flexible
manipulators. PhD thesis, Eindhoven University of
Technology, 1993.
[7] C. S. G. Lee. Robots kinematics, dynamics and
control. IEEE Computer, vol. 15(no. 12):6280,
1982.
[8] Bert van Beek. 3D measurement systems for robot
manipulators. Technical report WFW 98.011,
Eindhoven University of Technology, 1998.
[9] S. Wolfram. Mathematica, a System for Doing
Mathematics by Computer. Addison-Wesley
Publishing Company, 1993.