1

Status ofSmart Crane LabProject:Modeling and Control for a Forwarder Crane

Anton Shiriaev Leonid Freidovich Ian ManchesterUwe Mettin Pedro La Hera Simon Westerberg

Department of Applied Physics and ElectronicsUmea University, SE-901 87 Umea

E-mail: anton.shiriaev@tfe.umu.se

June 16, 2008

Abstract

The report suggests a short overview of the status of the project to June 2008. The project itself is focusedon developing kinematic and dynamic models for a crane commonly used in the forestry and mining applications.Based on these models we plan motions of a crane and suggest feedback controllers for their stabilization. Shortdescription of the system, challenges to control the crane,the performed experimental studies and of the proposedalgorithms is followed by the collection of copies of the papers published within the project in 2007-2008.

I. I NTRODUCTION AND SYSTEM DESCRIPTION

Vehicles equipped by cranes and that commonly used for cutting and picking-up timber littered on theground are depicted on Fig. 1. The reduced-size forwarder crane is shown on Fig. 2. As seen from the

Fig. 1. To the left: the photo of a forwarder (Komatsu Forest AB) picking-up a pile of timber. To the right: the schematic drawings offorestry vehicles: a forwarder and a harvester.

2

photos the crane is a large-size manipulator with open chainkinematics. Let us count a number of jointsof this manipulator:

• The first, second and third ones are revolute joints corresponding to kinematics of an elbow manip-ulator, see Fig. 3;

• The fourth joint is prismatic and corresponds to telescope of the crane;• At the end of the telescope joint, the crane has a gripper withthree degrees of freedom. Two degrees

of freedom (corresponding to rotation of the gripper aroundits axis of symmetry and closing-openingthe gripper) are actuated. The left degree of freedom is passive.

To summarize, the crane has seven joints with six of them actuated and one passive. Seven degrees offreedom are sufficient to describe configuration of the crane, if we disregard the position and orientation ofa mobile platform (i.e. a vehicle) it is mounted on. Taking into account the position and orientation of themobile platform, six new quantities (degrees of freedom) should be considered in addition to parameterizethe status of the crane relative to some inertia base frame.

The following parameters are identified for the crane depicted on Fig. 2

length [m] distance to CoG1 [m] mass [kg] inertia [kg m2]the 1st link 0.7 0.35 100 20the 2nd link 1.4 0.65 357.5 700.7the 3rd link 1.82 0.91 117.5 389.2the telescope 0.0, 1.55 0.82, 2.36 75 25

(min vs max extension) (min vs max extension)the gripper 100

Controlling the crane is the inherently difficult problem, while the control architecture irrespective ofits complexity has to have at least two levels:

• the first one for robust and efficient actuation of each joint of the crane (low level control),• the second one for generating commands to be performed by thecrane as a whole (high level control

and supervision) as parking, picking up a timber, Cartesian control (boom-tip control)etc.

Another layers (in addition to two mentioned) can be introduced as well for example for remote tele-operation over a communication channel (internet) to handle transmission of signals and visualizationof the crane status and its environment; or for the case of complete automation, where some cognitiveengines are required for motion generation.

Below we have listed a particular tasks for low and high levelsof control architecture, which areimportant and which were targeted in the project:

• Low level control design steps:– nonlinear friction identification,– angular position control of each link with feed-forward friction compensation,– hydraulic system modeling and identification,– parameters estimation of the mechanical plant,– model based control for the hydraulic system in under-damped links,– model based distributed control for each link angular motion control.

• High level control design steps:– design of the crane forward/inverse kinematics by the Denavit-Hartenberg convention,– record certain motions performed by professional drivers in order to automate specific tasks,– design smooth joint trajectories based on velocity profiles.

Some of these tasks are commented and elaborate below in the text. Later on, we briefly mention challengesand problems are to be approached and solved. At the end, the collection of the papers published withinthe project in 2007-2008 is added.

1CoG is the abbreviation for Centre of Gravity

3

II. N ONLINEAR FRICTION IDENTIFICATION AND CONTROLLERS FORINDIVIDUAL JOINTS

These tasks are the first since successful friction compensation and controlling angles of joints arenecessary even for identification of system dynamics, whichis unstable in open loop. Besides, carefultuning of controllers lead to a first step to automation.

A. Friction Identification

All mechanical systems are affected by friction, which is the nonlinearity in dynamics acting in theopposite direction to a speed of motion. The friction forces(or torques) appear in the crane due to physicalinteraction of rigid surfaces in contact (in a valve and in a joint shaft), due to viscosity in oil (fluids),due interaction between the oil and surface of pipeetc. The mechanical energy dissipated by friction isoften translated into a heat. It is common that if the friction effect is not taken into account for controldesign and is not compensated by an appropriate feed-forward signal, accuracy in the position control ofthe crane links cannot be achieve.

We identified a lumped model that represents the total friction present in the mechanical construction,hydraulic components and actuators. The friction is represented by a static map as a function of linkvelocity. The identified quantity here a feed-forward signal – ‘pseudo-friction,’ – since it is expressed asthe valve current required to overcome friction, rather than actual forces or torques.

The experiments for identifying this nonlinearity were organized and conducted in the following way:a step current is generated in the valve. It eventually results in a change of position of the correspondingactuator provided that the signal level lies above the stiction values. After some transient effect the velocityis approximately constant and it can be matched to the applied valve current. The magnitude of the stepsin the valve current are increased stepwise in order to produce a range of constant velocities beginningfrom zero.

As an example the resulting static friction maps for the firstlink is shown on Fig. 4. The static frictionaleffects are mainly caused by Coulomb friction, but can also have some contribution from dead zones inthe hydraulic valves. The identified maps can be used for compensation purposes in control design in astraightforward way since they are in terms of the input current.

B. Friction Compensation and PID Position Control

The classical approach for friction compensation is to use an estimate of the friction from velocitymapping identified previously, and add it with the opposite sign to the control signal. In Fig. 5 such acompensation scheme is illustrated as block diagram.

Conventionally PID (Proportional-Integral-Derivative) controllers have been tuned and used for thecontrolling each joint of the crane to achieve satisfactoryposition control. As a result of this tuningthe optimum performance has not not achieved, but at low velocities the performance of the crane weresatisfactory . An example of tracking a sinusoidal reference signal with the first link of the crane ispresented on Fig. 6.

III. H YDRAULIC SYSTEM MODELING AND MECHANICAL PARAMETER ESTIMATION

Having the system with a basic closed loop controller, procedures and methods for system identificationand parameter estimation can be applied in order to recover aparameterization the system dynamics ina accurate way. This step gives a better understanding of theprocess, separating the system as shown inFig. 7, for model based control design.

4

A. Identification of the Hydraulic Dynamics

Tracking some predefined position trajectories, the force delivered by the actuators are recorded andanalyzed for identification. The gray-box model for hydraulic force/torque generated by an input currentwith feedforward compensating the friction is the four order linear system with a transfer function,

τhydraulics =b3s

3 + b2s2 + b1s + b0

a4s4 + a3s3 + a2s2 + a1s + a0

(1)

Coefficients of numerator and denominator of (1) have been identified by various (direct and indirect)methods.

Key properties of the valve such as damping and natural frequency have been revealed in such analysis.The damping is a very important factor to know in advance, since in other links it would represent how thehydraulic dynamics would induce oscillations in the mechanical construction. They (oscillations) typicallyreduce the performance of trajectory tracking at high velocities. In Fig. 8 one example of measured forceis presented. Examine the plot, it is clear how oscillatory the response of the hydraulic system can be forcertain (step-like) input signals.

B. Mechanical System Parameter Estimation

Based on equations of motion (the second Newton law), and having the recorded input torque andangular position, it is possible to perform parameter estimation, in order to model the mechanical system.Considering that the system, from torque input to velocity asoutput, represents a first order model,

G =a0

s + b0

(2)

it is expected that

a0 ≈1

J, b0 ≈

b

J. (3)

unless other forces (torques), which appears, for example,due to imperfect friction compensation, flexibil-ity of links, uncertainty in mass distribution, i.e. possible mistakes in computing centers of masses of linksetc, are correctly computed and compensated. Tracking different trajectories and estimating velocities itis possible to identify such parameters through predictivearguments.

IV. CASCADE CONTROL

It is expected that by an appropriate control of the hydraulic force/torque, it is possible to counteractoscillations appeared in the cylinders in more efficient anddirect form where pressure measurements areavailable compare to the situation when the joint angles areonly measured.

Damping such oscillation can be approached based on variouscontrol techniques, we have chosenrobust and optimal control design to cope with that oscillations. In Fig. 9 a result of such a controllerfor one link is shown, where it is clearly seen the accurate tracking of the hydraulic system to a desiredtorque reference. Position control is achieved by model reference control.

A cascade control as shown in Fig. 10 is implemented for position control. The idea of this cascadecontrol scheme, is to use pressure measurements to compute the actual acting force/torque generated byeach actuator, and reduce oscillations while stabilizing the desired force and motion.

An example of the differences among single loop controller and cascade controller are shown in Fig. 11and Fig. 12. Fig. 11(a) shows the tracking of a link to some desired trajectory. In Fig. 11(b) the torque isshown, which clearly is not controlled and smooth, and at higher amplitudes and velocities would be moreoscillatory. On the other hand, Fig. 12(a) shows the tracking of the link to the same desired trajectory,but faster and with larger amplitude (trajectory that is notpossible with the previous controller). Fig.12(b) shows the torque which is under control action. This torque can be seen to be smoother and lessoscillatory than in the previous case.

5

V. FORWARD/INVERSEK INEMATICS AND TRAJECTORYGENERATION

Based on the diagram presented in Fig. 13, an analysis of the crane’s kinematics was performed inorder to calculate the 3D Cartesian orientation of the end effector.

Such analysis can be formulated asX = f(q) (4)

whereX represents the world coordinate which is a functionf(q) of the joint coordinatesqi . Differentinverse kinematics methods can be applied to exploit the redundancy property of the cranes mechanicaldesign, in order to find out the angular positionq having the world coordinateX .

The trajectories are designed in a way that a smooth bell-shaped velocity profile is predefined. This isdone by the use of the jerk as the optimisation criterion. Thejerk is the change in acceleration over time,the third time derivative of the position. For a given start time ts and end timete , the cost function tobe minimized is defined as:

C(θ) =1

2

∫ te

ts

(

d3θ

dt3

)2

dt (5)

resulting in that a function with a sixth derivative equal tozero (x(6) = 0) optimises the motion forminimum jerk. The general solution to this is a fifth order polynomial:

θ(t) = a0 + a1t + a2t2 + a3t

3 + a4t4 + a5t

5 (6)

By including known constraints to the motion, such as the motion duration, position, velocity andacceleration at start and end of the motion, a general solution can be found. It states the smoothestpossible transition from start to the end position.

VI. CHALLENGES AND STEPS TOBE DONE

Among others problems and tasks the next ones will be paid special attention:

• Developing algorithms for sensing and controlling crane under limited information. It isreasonable to assume that many of sensing devices will be of limited use in outdoor and fieldconditions. So that successful automation of crane operations will certainly depend on sensitivityand robustness of feedback controllers and motion plannersfor the case when some of easily availablemeasurements in laboratory conditions become unavailableor sensed with different resolution. Thisissue requires developing a method for reconstructing missing or uncertain signals (the so-calledsoft sensors or observers for missing variables) as well as developing control architectures that areincentive to large uncertainties in these signals.

• Developing vision based methods for estimating the status of the crane, the vehicle and itsenvironment. Cameras are one of few sensing devices that can be used for extracting informationfrom the scene in the front and from the back of the vehicle. This opens the possibility for char-acterization the status of the crane and the vehicle, enabling cameras to work as redundant sensingdevices for the vehicle. Redundancy in sensing is one of key requirements for fault detection andfault recovery in the system. This gives the opportunity forextraction and classification of items(logs, obstacles, humansetc) in the surrounding for the vehicle.

• Scenario based motion planning and control architectures.Not any motion of a crane will berequired to implement and there is no need to synthesize an universal controller able to cope witharbitrary disturbance or uncertainty in the system. Indeed, there are a few mostly important operational(working) cycles for the crane as well as particular expectation/request for human machine interfacefor crane operators. Elaborating and tuning control mechanisms for particular scenarios (as parking,Cartesian control) will be in focus.

6

Fig. 2. The photos of the reduced-size forwarder crane used for research experiments.

7

Fig. 3. The schematic drawing of an elbow manipulator, which representsthe first 3 joints of the crane on Fig. 2

−0.2 −0.1 0 0.1 0.2 0.3−0.3

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

0.4

0.5

pseu

do fr

ictio

n i f in

A

angular velocity ω in rad/s

friction identification − 1st link

α+0=0.355A

α+2=−0.7A/(rad/s)

α−0=−0.398A

α−2=−0.592A/(rad/s)

Fig. 4. Static pseudo-friction map for the first link.

e i theta2

est. i_f

est.omega2

theta2*

position

reference

control action

est. velocity

velocityestimation

velocity pseudo-friction

static mappseudo-friction

reference(desired position)

current position

real crane - 1st linkScope

error control action

PID Controller

Fig. 5. Friction Compensation block diagram as example for first link.

8

0 2 4 6 8 10 12 14 16 18 20−0.1

−0.05

0

0.05

0.1

0.15

time

angu

lar

posi

tion

desired trajectory vs ouput trajectory

Fig. 6. First Link measured position (green signal) vs First Link reference trajectory (blue signal).

Fig. 7. Identification-based modeling diagram.

0 0.5 1 1.5 2 2.5 3 3.5 4−4000

−2000

0

2000

4000

6000

8000

Time

y1

0 0.5 1 1.5 2 2.5 3 3.5 4−0.05

0

0.05

0.1

0.15u1

Time

Fig. 8. Input current and Force output for one experiment.

9

0 10 20 30 40 50 60−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

Fig. 9. Reference normalized torque and measured torque superimposed.

Hydraulic

Actuator

Mechanical

Device

Position

ReferencePosition

Position

Controller

Torque

Controller

Current Torque

Fig. 10. Two stages cascade control.

0 5 10 15 20 25 30 35 40−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

(a)0 5 10 15 20 25 30 35 40

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

0.4

0.5

(b)Fig. 11. (a) Reference signal and gripper position superimposed; (b) The resulting normalized torque.

10

0 2 4 6 8 10 12 14 16 18 20−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

(a)0 2 4 6 8 10 12 14 16 18 20

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

(b)Fig. 12. (a) Reference position trajectory and rotator angular position superimposed; (b) Hydraulic Torque under cascade feedback control.

d6

0,20

0,40

1,40

0,2

5

0,2

5

0,2

0

0,0

7

0,20

0,39

0,40

d1

d2

d3

r1

2

r2

r3

r4

r5

d4

d5

3

c1

c2

r6

d7 r

7

r0

Fig. 13. crane diagram.

Identification-based

Modeling and Control of

Hydraulically actuated Forestry Cranes

Pedro La Hera ∗ Andrej Zanhar ∗ Uwe Mettin ∗

Simon Westerberg ∗ Anton Shiriaev ∗,∗∗

∗ Department of Applied Physics and ElectronicsUmea University, SE-901 87, Umea, SWEDEN,

{Xavier.LaHera|Andrej.Zanhar|Uwe.Mettin|Simon.Westerberg|Anton.Shiriaev}@tfe.umu.se.∗∗ Department of Engineering Cybernetics

Norwegian University of Science and Technology, NO-7491 Trondheim,NORWAY.

Abstract: This article presents an example of experimental system identification basedmodeling and control carried out in an electro-hydraulic actuated crane. This developmentformulates an experimental approach to achieve distributed feedback control aiming at theautomation of different tasks for hydraulic actuated machines. The particular setup used inthis case is a crane of the type used on forestry vehicles known as forwarders, which travel off-road collecting logs cut by the harvesters. Different nonlinear dynamics are identified and usedlater for closed loop control. The control technique to be presented is able to asymptoticallytrack a reference trajectory despite model uncertainties in the mechanics and hydraulic systemdynamics. The performance is experimentally verified.

Keywords: Robotics in Agriculture and Forestry, Hydraulic Manipulator, SystemIdentification.

1. INTRODUCTION

The Swedish forest industry has a long-term goal of devel-oping autonomous and semi-autonomous forestry vehicles[3], [4].

There are mainly two types of off-road vehicle used in theforestry industry: the harvester, which fells and delimbsthe trees, and cuts the trunk into logs of a predeterminedsize, and the forwarder (see Fig. 1), which collects the logsin a tray and carries them to the nearest road for collection.

These two types of vehicle have similar on-board hydraulicmanipulators (cranes). An important stage towards au-tomation of forestry vehicles is establishing accurate dy-namical models of these manipulators, and designing high-performance low-level control systems. These can thenbe used in concert with high-level motion planners andteleoperation systems, e.g. [5].

Related recent publications include [7, 8, 9, 10]. Of partic-ular interest is the work of Munzer [11], which considersmany of the same issues on a crane with similar configu-ration to the one in our lab.

Unlike these methods, e.g [11], our objective is to studyan experimental approach to achieve the automation ofsuch a machine. Our foundation comes from the fact thatmodeling and control of a hydraulic system is far from⋆ This work is supported by the Center of Intelligent Off-roadVehicles (IFOR) at the Institute of Technology of Umea University,Komatsu Forest AB, Skogforsk, Sveaskog and the Kempe Founda-tion.

Fig. 1. A forwarder: the Komatsu 860.1.

trivial due to the highly nonlinear nature of the hydraulicsystem. As example, parameters such as bulk moduluschange drastically with oil temperature. The mechanicalconstruction is highly nonlinear itself, since parameters ofthe mechanical linkages may vary drastically and are usu-ally unknown (e.g. external payload, inertias). Moreover,significant uncertain nonlinearities such as external distur-bances, leakages and friction are unknown and cannot bemodeled accurately [14], and might change from season toseason, structural condition and age of the machine. Thesefactors result in significant difficulties to accurately modeland control a hydraulic system.

In this paper we present some experimental results workingtowards this goal. In order to be concrete with thispresentation, we will concentrate on the rotator of thegripper mounted at the boom tip (see Fig 4(b)). Thetask of the gripper is to position itself according to theorientation of the logs to pick them up, and move to thetray according to a predefined path without hitting thetray’s sides in accordance to the tips motion.

We emphasize the modeling and control of the end effectorsince it presents indirectly one more difficulty in contrastto the rest of the links, angular position sensing is not fea-sible with encoders. The manufacturing of this particularhydraulic motor does not allow to directly couple an en-coder in it. On the other hand, for an outdoor application,having an encoder mounted would not be the most robustway to sense its position, due to the rough environmentthis type of machines have to work on. Hitting the ground,trees, etc, would easily make an encoder to malfunction.For these reasons the solution adopted for this sensing willbe shortly described.

Even though results for only one link will be presented,the steps to be shown are applied to each link composingthe hydraulic manipulator.

The structure of the paper is as follows: in Section 2, theproblem formulation is posted defining some objectives;in Section 3, the experimental setup at Umea Universityis described in detail; in Section 4 steps on identificationof nonlinearities in the system is presented; in Section 5,we describe the method of friction compensation used;experimental identification of parameters of the mechani-cal setup as well as hydraulic system dynamics plants areinvestigated in Sections 6 and 7; exemplified control designtechniques for position feedback, torque feedback and acombination as cascade control are presented in Sections 8,9 and 10; some brief conclusions and discussions of futuredirections are given in Section 11.

2. PROBLEM FORMULATION ANDPRELIMINARIES

Equation of motion for any particular mechanical systemcan be formulated by the Newton’s second law

J · θ = τ (1)

where J is the inertia of the mechanical link, θ is the ro-tational motion expressed in radians and τ can informallybe thought of as the total external ”rotational forces” or”angular forces”. The external torque τ consists of variouscomponents which can be modeled as

τ = τhyd.dynamics − τfriction − τdisturbance (2)

where τhyd.dynamics is the hydraulic input torque thatcauses a change in rotational motion, τfriction andτdisturbance represent the friction and external distur-bances respectively.

If (2) is substituted into (1), it is obtained that dynamicsfor a particular link can be modelled by

J · θ = τhydraulics − τfriction − τdisturbance (3)

The friction torque τfriction is a nonlinear unknown map-ping acting in opposite direction to the angular velocity of

Fig. 2. Double acting piston-type actuating cylinder.

the link. It is often the case that the nature and the preciseform of that mapping is not needed, but its approximation,which can be used for (partly) compensating friction, is ofinterest. The classical approximation of the force of frictionbetween two solid surfaces is known as Coulomb frictionnamed after Charles-Augustin de Coulomb. According toCoulomb the frictional torque on each surface is exerted inthe direction opposite to its motion relative to the othersurface. On the other hand, Reynolds (1866) developedexpressions for the friction force caused by the viscosityof lubricants. The term viscous friction is used for thisfriction phenomenon. A friction model appears that iscommonly used in engineering [13]: the Coulomb plusviscous friction model. The friction torque can then bemodeled as

τfriction = b · θ + τCoulomb · sign(θ) (4)

where b is the viscous friction and τCoulomb represents theCoulomb part of the friction. Assuming that the system isfree of disturbances (e.g. noise), by substituting (4) into(3) it can be shown that

J · θ = τhydraulics − b · θ − τCoulomb · sign(θ) (5)

Assuming that τhydraulics has a direct relation with theinput current u, this expression could be rewritten in termsof the input current u instead of the hydraulic torque

J · θ = u − b · θ − τCoulomb · sign(θ) (6)

where J , b and τCoulomb will have a proportional relationto the real parameters J , b and τCoulomb. This differentialequation models the dynamics of a single link mechanicalstructure actuated by a hydraulic actuator. This expres-sion is valid under some assumptions. These assumptionscome mainly from the fact that the hydraulic functionshave complex nonlinear dynamics making τhydraulics beingnot exactly proportional to the input current u, but rathera dynamical response to it.

Since modeling the hydraulic force or torque from firstprinciples is far from trivial, due to the very large numberof parameters and nonlinearities, in this work it is shownhow to approximate the hydraulic torque response into alinear form, relating the input current (to the four-wayvalve) with the output force produced by the hydraulicactuator.

Let us consider a linear actuator for this purpose. Dis-regarding internal disturbances, nonlinearities caused byfriction, etc, the applied force produced by a linear hy-draulic actuator, as shown in Fig. 2, is given by [15]

F = AApA − ABpB (7)

where AA and AB are the areas at each respective cham-ber, pA and pB are the pressures measured at each cham-ber. The relations governing the dynamics of the pressuresare [15]:

pA = βVA

[−CempA − AAxp + qA]

pB = βVB

[−CempB + ABxp − qB ].(8)

where VA and VB are the chambers volumes, Cem is thecylinder internal leakage, xp is the pistons displacement,qA and qB are the input and out flow to and from thecylinder chambers. The linearized version of the hydraulicflow for qA and qB is [15]:

qA = 2Kqxs − 2Kc(pA − ps/2)qB = 2Kqxs + 2Kc(pB − ps/2)

(9)

where Kq and Kc are the valves coefficients and ps is thesupplied pump pressure and xs represents the motion ofthe four-way valve spool displacement.

The relation between the four-way valve spool position xs

and the input current u can be written as [16]

xs(s) =ω2

n

s2 + 2ξωn + ω2n

u(s) (10)

where ξ and ωn represent the damping ratio and naturalfrequency characteristics of the servo valve.

By combining equations (7) to (10) and by consideringthat the system parameters represent some numericalvalues, we obtain a linear model, in which parametershave been collected and substituted to simplify notation,defined by the transfer function:

Fcyl =AA·cω2

n(s+f)+AB ·hω2

n(s+a)(s+a)(s+f)(s2+2ξωn+ω2

n) · u(s)

+−AA·b(s+f)−AB ·g(s+a)(s+a)(s+f) xp

+AAe(s+f)−ABm(s+a)(s+a)(s+f)

(11)

Based on this analysis, it is shown that dynamics of thehydraulic force/torque can be represented by a fourthorder linear time invariant system, i.e 11.

The objectives to be covered along this paper can besummarized as follows,

• experimental procedure for friction identification,• proposing a friction compensation strategy for this

type of setup,• parameter estimation based on the current-position

dynamics represented by (6),• parameter estimation based on the current-hydraulic

torque-position dynamics represented by (11)-(5),• designing of a closed loop strategy based on position

measurements only with current being the inputsignal,

• designing of a closed loop strategy based on pressureand position measurements considering the currentbeing the input signal to the hydraulic system andthis one to the mechanical system (cascade control),

• a comparison of the performance of these two con-trollers.

3. EXPERIMENTAL SETUP

Experimentations and tests are carried out at the SmartCrane Lab located at Umea University. The Labora-tory is equipped with an electro-hydraulic actuated craneCRANAB model 370RCR (see Fig. 3), a forwarder cranewhich is somewhat smaller than most cranes on productionforwarders, but similar in configuration and dynamics.

Fig. 3. Crane installed at the Department of AppliedPhysics and Electronics, Umea University.

The hydraulic hardware in the Smart Crane Lab consistsof:

• hydraulic cylinders manufactured by Valmet,• hydraulic motor for the rotator at the gripper manu-

factured by Valmet,• a unit containing six servo-valves from Sauer-Danfoss

model L90LS,• the power supply for this system consists of an

electrical motor driving a hydraulic pump (type H4-010214-132) set to provide a constant supply pressureof 180 bars for the whole system operation.

In addition, the associated sensing equipment includes:

• encoders of 4000 pulses/turn, present to measure thevarious links angular positions,

• pressure transducers (HD 3403-10-C3.39) capable ofsensing in a range of [0, 200] bar.

The crane can be directly manipulated by a chair, sameas the ones mounted in the cabin of real forwarders. Thischair contains buttons and joysticks that allow the driversto have full control over the whole machine operationand the crane. Signals from this chair are handled by theprocessing unit.

For the control of the crane as well as the implementationof algorithms a dSPACE Prototyping Hardware is used.The processing unit applied in this particular case isthe MicroAutoBox (MABX), which directly controls theavailable I/O features, such as the Electronic ControlUnit, the AD and DA converter units, the digital I/O andCAN subsystems. The control models are implemented asexecutable code on the PPC.

In order to provide a sufficient range of current to drive theservo valves a RapidPro unit is installed. The RapidProcontains a so-called Power Unit (PU) which transformsthe low voltages generated by the MABX to appropriate

currents for the valve solenoids. The current in each circuitcan be measured. Furthermore, there is a so-called SignalConditioning Unit (SCU) which can handle the incomingmeasurement signals to voltage levels needed to be fed tothe MABX. Finally, a Dell PC is present to monitor andserve as an on-line user interface through the use of ControlDesk.

3.1 Gripper’s rotator sensing device

This is shortly the solution adopted for the sensing ofthe rotator’s angular position. A metallic disk is attachedat the rotator as shown in Fig. 4(b). Along the disk atotal of thirty-six magnets are located to have a resolutionof ten degrees as seen in Fig. 4(a). A magnetic sensormanufactured by IFM electronics is used to sense thesemagnets. The signals coming from the magnetic sensorsare read and converted to position values by the processingunit.

(a) (b)Fig. 4. (a) CAD model of the disk designed for the rotator’s

angular position sensing; (b) Frontal view of theinstallation of such a device.

4. IDENTIFICATION OF NONLINEARITIES

Forestry machines are constructed in a robust way tohandle different terrains and climates. For control purposesthis type of system becomes a highly complicated nonlin-ear system with different uncertain nonlinear parametersthat need to be identified throughout designed tests.

For instance, the valve has a significant dead-zone due tooverlap on the spool in order to avoid leakages. Saturationto its maximum actuation level is also present. Moreover,in such heavy duty machine, friction occurs in the mechan-ical structure as well as in the hydraulic components andactuators, these effects have to be taken into account.

In order to have a reliable compensation for nonlinearitiesmany tests were performed in order to extract all thepossible information.

4.1 Identification of friction

A lumped model representing the total friction (4) presentin the mechanical construction, hydraulic components andactuators, was identified. The friction is modeled by astatic map as a function of current values with respect tolink velocity [13]. The expression “pseudo-friction” is used,since it represents the valve current required to overcomefriction, rather than actual forces or torques.

The experiments were conducted as follows: a step currentis generated in the valve which eventually results in achange of position of the corresponding actuator providedthat the signal level lies above the stiction values. Aftersome transient effect the velocity reaches a constant steadystate and it can be matched to the applied valve current.The magnitude of the steps in the valve current areincreased stepwise in order to produce a range of constantvelocities beginning from zero.

As example in Fig. 5 the resulting static friction maps forthe gripper’s rotator is shown. The static frictional effectsare mainly caused by coulomb and viscous friction, butcan also have some contribution from dead zones in thehydraulic valves. The identified maps can be used for com-pensation purposes in control design in a straightforwardway since they are in terms of the input current to thefour-way valve.

−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2−0.7

−0.6

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

angular velocity [rad/s]

stat

ic p

seud

o−fr

ictio

n (n

orm

aliz

ed c

urre

nt)

Fig. 5. Static pseudo-friction map for the first link.

5. FRICTION COMPENSATION

In order to have high accuracy in the position control ofthe crane links, some undesired effects mainly caused byfriction need to be removed.

The classical approach is to use a feed forward term addedto the control signal using an estimate of the frictionidentified previously, derived from an estimate of thelink velocity. In Fig. 6 such a compensation scheme isillustrated as block diagram.

e i theta2

est. i_f

est.omega2

theta2*

position

reference

control action

est. velocity

velocityestimation

velocity pseudo-friction

static mappseudo-friction

reference(desired position)

current position

real crane - 1st linkScope

error control action

PID Controller

Fig. 6. Friction Compensation block diagram as examplefor first link.

However, if there is very large component of Coulombfriction, as in our case, then noisy estimates of link velocityaround zero can lead to a “chattering” effect [13]. A furtherproblem is that, if the link is stationary and a small control

input has been applied, the friction compensator will applyany compensation signal, since the link velocity is zero.

To overcome these difficulties, we apply a modification inthe classical friction compensation approach, which in ourparticular case gives good results in the practical sense.The velocity-estimate used in the compensation signal iscalculated like so:

ˆθ = (1 − α(r))θ + rα(r) + µ sign(i). (12)

Here, θ is the velocity estimated from a simple differ-entiator and low-pass filter arrangement, and r is thederivative of the reference signal, which is known exactly.The function α(r) is a bell-shaped function which is equalto one at r = 0, and smoothly drops to zero outside therange [−0.1, 0.1] rad/s. The motivation is that, aroundzero, assuming reasonably good tracking, the referencevelocity is smoother and more reliable measure of velocitythan a noisy estimate of the link velocity.

The term µ sign(i) is added to overcome the second prob-lem described above. If the link is stationary, then thedirection of the valve current i is used to decide to whichdirection the friction compensation force should be ap-plied. The constant µ is chosen to be a small value, suchas 0.001.

6. MECHANICAL PLANT MODEL PARAMETERESTIMATION

Applying the friction compensation proposed, to compen-sate Coulomb friction only, that is

u = v + τCoulomb · sign(θ),

where v is the nominal input signal, removes strongnonlinearity in the system. Relations given by (6) and (4)can be written as

J · θ =(

v + τCoulomb · sign(θ))

− b · θ − τCoulomb · sign(θ)

= v − b · θ + e(t), (13)

where parameters J , b are to be identified, and e(t) can beinterpreted as the modeling error.

The system (13) is unstable in respect to the position θ(t)(the rotator moves in horizontal plane and its potentialenergy is constant). It is suggested

• to introduce a feedback action to stabilize the system;• to use a reference signal to the closed loop signal of a

particular band-pass characteristic.

The simplest choice of a stabilizing controller for (13) isthe proportional feedback

v(t) = Kp ·

(

r(t) − θ(t))

where r(t) is the reference and Kp the proportional gain.With such a choice the system (13) becomes

J · θ + b · θ + Kp · θ = Kp · r(t) + e(t). (14)

and is stable provided that Kp > 0. By analyzing theLaplace Transform of (14) as

J · s2θ(s) + b · sθ(s) + Kp · θ(s) = Kp · r(s) + e(s). (15)

a second order transfer function given by

θ(s)

r(s)=

Kp/J

s2 + b/Js + Kp/J, (16)

relates the reference signal r(t) with the angular positionθ(t). This transfer function opens the possibility to identifya parametric model described by

G(s) =b1s + b0

s2 + a1s + a0, (17)

having b1 ≈ 0, b0 ≈ Kp/J , a1 ≈ b/J , a0 ≈ Kp/J .

Various experiments varying Kp were performed and thereferences r(t) were built with sinusoids varying within thefrequency range [2, 6] rad/sec. For the case when the gainKp is chosen to be 0.04, it is obtained that

G(s)Kp=0.04 =0.001262s + 2.526

s2 + 4.589s + 3.543, (18)

and for the case when Kp = 0.16,

G(s)Kp=0.16 =0.01319s + 26.42

s2 + 10.05s + 25.24, (19)

Fig. 7 shows the recorded signals and the validation testsfor these two cases. The averaged values taken from thisestimation gives

¯J = 0.0085,

¯b = 0.0554.

0 2 4 6 8 10 12 14 16 18 20−2.5

−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

2.5

(a)0 5 10 15 20 25

−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

y1

Time

Measured Output and Simulated Model Output

Measured Outputm1 Fit: 85.37%

(b)

0 2 4 6 8 10 12 14 16 18 20−2.5

−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

2.5

(c)0 5 10 15 20 25

−2.5

−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

y1

Time

Measured Output and Simulated Model Output

Measured Outputm1 Fit: 85.12%

(d)Fig. 7. Identification data for two particular values of

the gain Kp: (a) Recorded reference signal r(t) andangular position θ(t) as functions of time for the casewhen Kp = 0.04; (b) identified plant validation plotfor the case when Kp = 0.04; (c) recorded referencesignal r(t) and angular position θ(t) as functions oftime for the case when Kp = 0.16; (d) identified plantvalidation plot for the case when Kp = 0.16.

7. HYDRAULIC PLANT MODEL IDENTIFICATION

According to the analysis presented in section 2, thehydraulic force/torque can be modeled by a fourth ordertransfer function

τhydraulics =b3s

3 + b2s2 + b1s + b0

a4s4 + a3s3 + a2s2 + a1s + a0(20)

The input signal in this case consists of a step signal withadditional random disturbances to induce the hydraulicsto reveal the valve dynamics (see Fig 8(a)). An analysisof the spectral density (see Fig 8(b)) reveals the secondorder properties of the valve given by (10). Differentidentification methods were applied (see Fig 8(c)), amongof the best results are the predictive error method (PEM)and the output error method (OE). The transfer functionobtained is

20.79s3− 1696s2 + 8.005e004s + 1.139e005

s4 + 35.68s3 + 1428s2 + 1.972e004s + 2.783e004(21)

which indirectly implies that the valve operates around afrequency of approximately 32 rad/sec (see Fig 8(d)), andthe damping of the system is rather small. The dampingis a very important factor to know in advance, since inother links it would represent how the hydraulic dynamicswould induce oscillations in the mechanical constructionreducing the performance of trajectory tracking at highvelocities.

0 1 2 3 4 5 6−0.2

0

0.2

0.4

0.6

0.8

Time

y1

0 1 2 3 4 5 6−0.05

0

0.05

0.1

0.15

0.2u1

Time

(a)

10−1

100

101

102

10−1

100

101

Am

plitu

de

From u1 to y1

10−1

100

101

102

−800

−600

−400

−200

0

Pha

se (

degr

ees)

Frequency (rad/s)

g_spa15g_spa25g_spa50

g_spa15g_spa25g_spa50

(b)

0 1 2 3 4 5 6−0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

y1

Time

Measured Output and Simulated Model Output

Measured Outputoe341 Fit: 76.46%m2c Fit: 81.3%m4c Fit: 77.33%m4o Fit: 79.13%pss5 Fit: 75.29%n4s4 Fit: 77.32%m3c Fit: 76.42%

(c)

10−1

100

101

102

103

104

10−3

10−2

10−1

100

101

Am

plitu

de

From u1 to y1

10−1

100

101

102

103

104

−500

−400

−300

−200

−100

0

Pha

se (

degr

ees)

Frequency (rad/)

(d)Fig. 8. Identification data: (a) Recorded input signal

u(t) and normalized differential of pressure response(torque); (b) spectral density analysis of data; (c)results of different identification methods; (d) Bodediagram and spectral density superimposed.

8. ANGULAR POSITION CONTROL DESIGN

Despite a large volume of published work on advancedcontrol strategies, modern industrial robots typically em-ploy control systems that are independent for each jointand thus treat each joint and its associated drive as asimple servo mechanism. These joint servos conventionallyuse PID (Proportional, Integral and Derivative) controlwhich regards the complicated effects of changing iner-tia, centripetal and Coriolis terms, gravity and frictionaltorques as disturbances and makes no attempt to explicitlycompensate for them. As a result optimum performanceis frequently not achieved [2]. As an alternative to PID,there are two basic control techniques which attempt toreduce the undesirable effects of coupled torques/forcesand changing inertia. These are Model Reference Control(MRC) (see e.g. [1]), which uses a mathematical modelof the robot to explicitly compensate for dynamic terms,

Fig. 9. Block Diagram of the model reference control(MRC) structure

and adaptive control (see e.g. [14]) which automaticallytunes the parameters in the control system on the basis ofobservations of the robot behavior.

Each control technique has advantages and disadvantages,MRC is relatively easy to implement and gives good resultswhen tested in simulation. However it can be difficult toobtain precise dynamic models of real robots, particularlywhen significant Coulomb or slip/stick Coulomb frictionexists.

The majority of published work on robot control systemshas been based on either theoretical study of control sys-tem behavior or numerical experiments conducted throughcomputer simulation. It is unfortunately the case thatMRC is not well suited to investigation through simula-tion. The MRC controller requires a mathematical modelof the system dynamics. Similarly, the simulation requiresa model of the system dynamics. It is clear that the use ofthe same model in both the controller and the simulationcan result in apparently perfect controller performance. Itis the present authors’ belief that the benefits of MRC canonly be properly demonstrated by controlling a physicalsystem, rather than a simulation. In the following lines theuse of this method based on identified models describedabove will be exploited for position control.

The proposed closed loop scheme that will be used isshown in Fig. 9. The PID gains can be tuned by differentmethods, e.g Ziegler Nichols, optimization based, etc. It isseen from Fig. 9 that the PID named as A will be drivenby the process error. This particular solution enables thecontrollers A and B to interact among each other for effectsproduced by the system disturbance z. The model appliedfor such a controller was experimentally found as explainedin section 6; with a similar analysis it is possible to finda reference model for each link. Friction compensation isapplied as presented in section 4.

A plot showing the performance of this controller togetherwith a feed-forward friction compensation is shown inFig. 10. Fig. 10(b) shows the response of the hydraulicsystem (not controlled at this stage yet) to this particularcontrol scheme. Although the tracking to the referencesignal shows to be somewhat accurate, taking into accountthat a resolution of ten degrees is achieved by the magneticsensor, the hydraulic torque shows an oscillatory behavior.For the case of the gripper such hydraulic oscillations arenot dramatic, but in the case of the rest of the links it isunderstood that they could induce mechanical oscillations

provoking a lose of efficiency while tracking referencesignals.

0 5 10 15 20 25 30 35 40−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

(a)0 5 10 15 20 25 30 35 40

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

0.4

0.5

(b)Fig. 10. (a) Reference signal and gripper position super-

imposed; (b) The resulting normalized torque.

9. PRESSURE CONTROL DESIGN

Having a SISO linear system represented by (21) opensthe possibility to apply linear control theory to accuratelycontrol the hydraulic torque. This control is based on thefact that hydraulic functions are not decoupled, and sodistributed position control of each link would not give anoptimum performance when moving at high velocities. Byhydraulic torque/force control, smoothness during trajec-tory tracking is planed to be achieved.

A rather simple PID with friction compensation wasimplemented. The PID gains are found in this case byoptimization procedure. The optimization criteria appliedis based on integrating the square of the system error(IAE) over a fixed interval of time. Parameters found byapplying this optimization procedure are

Kphyd = 0.2, Kihyd = 5.6, Kdhyd = 0.02.

Fig. 11 shows the tracking of the hydraulic torque ac-cording to some reference signal which is composed by asummatory of sinusoidal signals of different frequencies. Itcan be seen that accurate torque control can be achievedby applying feedback control designed with the argumentspresented in Section 7. In Fig. 12 the input signal gener-ated from the controller is shown. What is important tonotice is how the friction compensation behaves along thetracking.

0 10 20 30 40 50 60−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

Fig. 11. Reference Signal and normalized torque superim-posed

0 10 20 30 40 50 60−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

Fig. 12. Input current to the four way valve

10. CASCADE CONTROL ALGORITHM

It is expected that by an appropriate control of the hy-draulic force/torque, it is possible to counteract oscilla-tions much more directly than via measurements of angu-lar positions. In order to combine the design techniquespresented in Sections 8 and 9 a cascade control as shownin Fig. 13 is thought to be implemented. The idea of thiscascade control scheme, is to use pressure measurementsto compute the actual acting force/torque generated byeach actuator, and reduce oscillations while stabilizing thedesired force and motion.

Hydraulic

Actuator

Mechanical

Device

Position

ReferencePosition

Position

Controller

Torque

Controller

Current Torque

Fig. 13. Two stages cascade control.

The control problem in this sense is split into two stages.The outer loop controller calculates the reference pistonforces Fref needed to drive the manipulator along thepredefined joint space θref . The inner loop controller takesthis force reference Fref and computes the servo-valveinput current u needed to make the true piston forces Fasymptotically track Fref . Since F asymptotically tracksFref , which is itself an asymptotically stabilizing controlfor the manipulator motion around θref , the overall cas-cade system is asymptotically stable around the trajectoryreference θref . The same idea is applied in the angular casefor the rotator and slewing.

In Fig. 14 a plot showing a reference position trajectoryand the recorded angular position from the rotator ispresented. Tracking this particular trajectory is of betterquality as in Fig. 10(a). Unlike the case of Fig. 10(b), thetorque acting under cascade control (see Fig. 14(b)) allowsus to properly control it during the path tracking, and so itis expected that the angular position will behave smootherwhen pressure oscillations are diminished.

0 2 4 6 8 10 12 14 16 18 20−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

(a)0 2 4 6 8 10 12 14 16 18 20

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

(b)Fig. 14. (a) Reference position trajectory and rotator

angular position superimposed; (b) Hydraulic Torqueunder cascade feedback control.

11. CONCLUSIONS AND FUTURE WORK

The primary target of this project is to increase produc-tion efficiency of forestry machines by means of intelli-gent control techniques. Essential to any autonomous orsemi-autonomous forestry crane is good low-level systemidentification and control. In this paper we have presentedexamples of results along these lines. The main problemsto overcome were related to friction identification andcompensation, system modeling based on experimentalidentification, and control design of different levels.

There is much that can be done to continue this project. Atpresent the cascade-control structure, which showed verypromising results in damping of oscillations, has only beenimplemented on some links of the crane. Extending this toall links is a natural evolution of the current work, and itis expected that this will allow much smoother trajectorytracking to be achieved.

A fully experimental approach for modeling the mechan-ical system as well as the hydraulic dynamics was pre-sented. A question that arises in an industrial point ofview is if it would be possible to control the crane by onlypressure sensing. This cannot be answered yet, but an ideaof using different techniques of observer based controllersaccording to identified plants and sensor fusion is beingstudy at the moment.

The identification procedure presented opens the possibil-ity to apply different type of control design approaches. Itis indirectly the aim of the project to achieve transporta-bility of the design technique to cranes of different scales.

Other related work includes using the proposed low-level in concert with higher-level control schemes, suchas semi-autonomous tele-operation systems (see, e.g., [5]),and fully-autonomous systems based on optimized motionplanning.

REFERENCES

[1] B. Siciliano and L. Villani. “Robot Force Control”.Kluwer Academic Publishers, 1999

[2] Craig, J. J., “Introduction to Robotics: Mechanicsand Control”, Prentice Hall, 2003

[3] U. Hallonborg, “Forarlosa skogsmaskiner kan blilonsamma (Unmanned forestry machines can be com-petitive)” (in Swedish), Skogforsk Results, No 9, 2003.

[4] M. Brander, D. Eriksson, B. Lofgren, “Automation ofknuckleboom work can increase productivity”, Skog-forsk Results, No 4, 2004.

[5] S. Westerberg, I. R. Manchester, U. Mettin, P.La Hera, A. Shiriaev, “Virtual Environment Tele-

operation of a Hydraulic Forestry Crane”, submit-ted to the 2008 IEEE International Conferenceon Robotics and Automation, available online atwww.tfe.umu.se/forskning/Control Systems/teleoperation icra.pdf .

[6] P. D. Lawrence, R. V. Ross, Articulated arm controlIn: US Pat. Nr. 5,062,755, 1991.

[7] P. D. Lawrence, , F. Sassani, B. Sauder, N. Sepehri,U. Wallersteiner, J. Wilson, “Computer-assisted con-trol of excavator-based machines”, International Off-Highway and Powerplant Congress and Exposition,Milwaukee, Wisconsin, 1993.

[8] J. Mattila, T. Virvalo (a), “Computed force controlof hydraulic manipulators”, 5th Scandinavian Inter.Conf. On Fluid Power pp. 139154, 1997.

[9] M. Linjama, The Modelling and Actuator Space Con-trol of Flexibile Hydraulic Cranes, PhD thesis, Tam-pere University of Technology, Tampere, Finland,1998.

[10] M. E. Munzer, P. Pedersen, “Real time simulationmodel of flexible mobile crane for machine-operatorinteraction testing” Proc. 2nd International Work-shop on Computer Software for Design, Analysis, andControl of Fluid Power Systems, 2001.

[11] M. E. Munzer, Resolved Motion Control of MobileHydraulic Cranes, PhD Thesis, Aalborg University,Denmark, 2002.

[12] Graham C. Goodwin, Stefan F. Graebe, MarioE. Salgado, Control System Design, Prentice Hall,New Jersey, 2001.

[13] H. Olsson, K. J. Astrom, C. Canudas de Wit,M.Gafvert & P. Lischinsky, “Friction Modelsand Friction Compensation”. European Journal ofControl, Dec. 1998, no.4, pp. 176-195.

[14] F. Bu and Bin Yao, “Adaptive robust precision mo-tion control of single-rod hydraulic actuators withtime varying unknown inertia: a case study”, In Proc.of American Control Conference, 4129-4133, 2000.

[15] Noah D.Manring. Hydraulic Control Systems. JohnWiley & Sons, New York, 2005.

[16] J.A.F. Ferreira, Modelacao de Sistemas Hidraulicospara Simulacao com Hardware-in-the-loop. PhD The-sis (in Spanish), University of Aveiro, Portugal, 2003.

[17] Kemin Zhou, John C. Doyle, Essentials Of RobustControl, Prentice Hall, 1997.

[18] I. R. Petersen, V. A. Ugrinovski, and A. V. Savkin,Robust Control Design using H∞ Methods, Springer-Verlag, London, 2000.

Identification and Control of a

Hydraulic Forestry Crane

Pedro La Hera ∗ Uwe Mettin ∗ Ian R. Manchester ∗

Anton Shiriaev ∗,∗∗

∗ Department of Applied Physics and ElectronicsUmea University, SE-901 87, Umea, SWEDEN,

{Xavier.LaHera|Uwe.Mettin|Ian.Manchester|Anton.Shiriaev}@tfe.umu.se.∗∗ Department of Engineering Cybernetics

Norwegian University of Science and Technology, NO-7491 Trondheim,NORWAY.

Abstract: This article presents the identification and control of an electro-hydraulic crane.The crane is of the type used on forestry vehicles known as forwarders, which travel off-roadcollecting logs cut by the harvesters. The dynamics identified include significant frictional forces,dead zones, and structural and hydraulic vibrations. The control algorithm proposed, comprisedof a linear controller and a compensator for nonlinearities, is able to accurately track a referencetrajectory for the end effector, despite uncertainties in the arm mechanics and hydraulic systemdynamics. A further control design is presented which uses an inner loop to compensate forvibrations in the hydraulic system, and its performance is experimentally verified.

Keywords: Robotics in Agriculture and Forestry, Hydraulic Manipulator, SystemIdentification.

1. INTRODUCTION

The Swedish forest industry has a long-term goal of devel-oping autonomous and semi-autonomous forestry vehicles[1], [2].

There are mainly two types of off-road vehicle used in theforestry industry: the harvester, which fells and delimbsthe trees, and cuts the trunk into logs of a predeterminedsize, and the forwarder, which collects the logs in a tray,and carries them to the nearest road for collection. Aforwarder is shown in Fig. 1.

These two types of vehicle have similar on-board hydraulicmanipulators (cranes). An important stage towards au-tomation of forestry vehicles is establishing accurate dy-namical models of these manipulators, and designing high-performance low-level control systems. These can thenbe used in concert with high-level motion planners andteleoperation systems, e.g. [3].

Related recent publications include [4, 5, 6, 8]. Of particu-lar interest is the work of Munzer [7], which considers manyof the same issues on a crane with similar configuration tothe one in our lab.

In this paper recent experimental results on a real for-warder crane working towards this goal will be presented.Each joint controller uses an angular position sensor (en-coder) for feedback control of the joint position. An inner-loop control based on direct control of cylinders pres-

⋆ This work has been supported by the Center of Intelligent Off-road Vehicles (IFOR) at the Institute of Technology of UmeaUniversity, Komatsu Forest AB, Skogforsk, Sveaskog and the KempeFoundation.

Fig. 1. A forwarder: the Komatsu 860.1.

sure/force has been implemented for the first link of thecrane, and shows impressive results on damping of oscil-lations. However, it is not yet implemented on the otherlinks.

The structure of the paper is as follows: in Section 2, adetailed description of the experimental setup is described;in Section 3 identification of nonlinearities in the systemis presented; in Section 4 the method of friction compen-sation used is explained; experimental results for trackinga reference trajectory are given in Section 5; in Section6 some early results on direct control of hydraulic forceusing pressure sensors is given; some brief conclusions anddiscussions of future directions are given in Section 7.

Fig. 2. Crane installed at the Department of AppliedPhysics and Electronics, Umea University.

2. EXPERIMENTAL SETUP

Experimentations and tests are carried out at the SmartCrane Lab located at Umea University. The Labora-tory is equipped with an electro-hydraulic actuated craneCRANAB model 370RCR (see Fig. 2), a forwarder cranewhich is somewhat smaller than most on cranes on produc-tion forwarders, but similar in configuration and dynamics.

The hydraulic hardware in the Smart Crane Lab consistsof:

• hydraulic cylinders manufactured by Valmet,• a unit containing six servo-valves from Sauer-Danfoss

model L90LS,• the power supply for this system consists of an

electrical motor driving a hydraulic pump (type H4-010214-132) set to provide a constant supply pressureof 180 bars for the whole system operation.

In addition, the associated sensing equipment includes:

• encoders of 4000 pulses/turn, present to measure thevarious links angular positions,

• pressure transducers (HD 3403-10-C3.39) capable ofsensing in a range of [0, 200] bar.

The crane can be directly manipulated by a chair, sameas the ones mounted in the cabin of real forwarders. Thischair contains buttons and joysticks that allow the driversto have full control over the whole machine operationand the crane. Signals from this chair are handled by theprocessing unit.

The processing unit applied in this particular case is thedSPACE MicroAutoBox (MABX), which directly controlsthe available I/O features, such as the Electronic Con-trol Unit, the AD and DA converter units, the digitalI/O and CAN subsystems. MATLAB/Simulink is usedto implement executable code for the processing unit. Inorder to provide a sufficient range of current to drive theservo valves a RapidPro unit is installed. The RapidProcontains a Power Unit (PU) which transforms the lowvoltages generated by the MABX to appropriate currentsfor the valve solenoids. The current in each circuit canbe measured. Furthermore, there is a Signal ConditioningUnit (SCU) which can handle the incoming measurementsignals to voltage levels needed to be fed to the MABX.

Finally, a Dell PC is present to monitor and serve as anon-line user interface through the use of Control Desk.

3. IDENTIFICATION OF NONLINEARITIES

Forestry machines are constructed in a robust way tohandle different terrains and climates. For control purposesthis type of system becomes a highly nonlinear systemwith uncertain parameters that need to be identifiedexperimentally.

For instance, the valve has a significant dead-zone due tooverlap on the spool in order to avoid leakages. Saturationto its maximum actuation level is also present. Moreover,in such heavy duty machines, friction occurs in the me-chanical structure as well as in the hydraulic componentsand actuators, effects that have to be taken into account.

In order to have a reliable compensation for nonlinearitiesmany tests were performed in order to extract all thepossible information.

3.1 Identification of friction

A lumped model representing the total friction, present inthe mechanical construction, hydraulic components andactuators, was identified. The friction is modeled by astatic map as a function of current levels with respect tolink velocity [9]. The expression “pseudo-friction” is used,since it represents the valve current required to overcomefriction, rather than actual forces or torques.

The experiments were conducted as follows: a step currentis generated in the valve which eventually results in achange of position of the corresponding actuator providedthat the signal level lies above the stiction values. Aftersome transient effect the velocity reaches a constant steadystate and it can be matched to the applied valve current.The magnitude of the steps in the valve current areincreased stepwise in order to produce a range of constantvelocities beginning from zero.

For example, in Fig. 3 the resulting static friction mapsfor the first link is shown. The static frictional effectsare mainly caused by Coulomb and viscous friction, butcan also have some contribution from dead zones in thehydraulic valves. The identified maps can be used for com-pensation purposes in control design in a straightforwardway since they are in terms of the input current to thefour-way valve.

4. FRICTION COMPENSATION

In order to have high accuracy in the position control ofthe crane links, some undesired effects mainly caused byfriction need to be removed. The classical approach is touse a feed-forward term added to the control signal usingan estimate of the friction derived from an estimate of thelink velocity [9]. In Fig. 4 such a compensation scheme isillustrated as block diagram.

However, if there is very large component of Coulombfriction, as in our case, then noisy estimates of link velocityaround zero can lead to a “chattering” effect [9]. A furtherproblem is that, if the link is stationary but a small

−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2−0.7

−0.6

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

angular velocity [rad/s]

stat

ic p

seud

o−fr

ictio

n (n

orm

aliz

ed c

urre

nt)

Fig. 3. Static pseudo-friction map for the first link.

e i theta2

est. i_f

est.omega2

theta2*

position

reference

control action

est. velocity

velocityestimation

velocity pseudo-friction

static mappseudo-friction

reference(desired position)

current position

real crane - 1st linkScope

error control action

PID Controller

Fig. 4. Friction Compensation block diagram as examplefor first link.

control input has been applied, the friction compensatorwill randomly apply any compensation signal, since thelink velocity is zero.

To overcome these difficulties, a modification in the classi-cal friction compensation approach is proposed, which inour particular case gives satisfactory results in the practi-cal sense. The velocity-estimate used in the compensationsignal is calculated like so:

ˆθ = (1 − α(r))θ + rα(r) + µ sign(i). (1)

Here, θ is the velocity estimated from a differentiator andlow-pass filter arrangement, and r is the derivative of thelink’s position reference signal, which is known exactly.The function α(r) is a bell-shaped function which is equalto 1 at r = 0, and smoothly drops to zero outside therange [−0.1, 0.1] rad/s. The motivation is that, aroundzero, assuming reasonably good tracking, the referencevelocity is smoother and more reliable measurement ofvelocity than a noisy estimate of the link velocity.

The term µ sign(i) is added to overcome the second prob-lem described above. If the link is stationary, then thedirection of the valve current i is used to decide to whichdirection the friction compensation force should be ap-plied. The constant µ is chosen to be a small value, suchas 0.001.

5. REFERENCE TRAJECTORY TRACKINGEXPERIMENTS

In this section, results of a trajectory tracking experimentis shown. The control strategy is based on PID positionfeedback control and a feed-forward term for friction

−1 0 1 2 3 4−2

−1

0

1

2

x−position [m]

z−po

sitio

n [m

]

autonomous motion − circle

Fig. 5. Desired reference trajectory for the boom-tip intwo-dimensional space. Period of rotation: 25 seconds

compensation (see Fig. 4). The intention is not to showa sophisticated motion planner, but rather evaluate theperformance of the reference tracking utilizing positionfeedback and friction compensation.

In Fig. 5 the desired motion is depicted. Here, the boom-tip has to follow a circular reference counter-clockwise ina two-dimensional space. Although not a trajectory thatwould be common in practice, it has many regions closeto zero link velocity, and thus provides a useful test of thefriction compensation presented above.

By inverse kinematics it is possible to compute individualtrajectories for each joint if a particular trajectory forthe telescopic arm is defined. Fig. 6 shows the generatedtrajectory and the final crane motion super-imposed. Asit can be seen, the reference tracking is quite accurate,however there exist some oscillations during the liftingphase.

The error for such a trajectory is less than 15 mm inaverage. During the cycle the end effector is moving atapproximately 0.2m/s. Oscillations show up particularlyduring the crane’s lifting phase, observable for the firstquadrant of the phase angle. This can be explained bythe fact that this is a hydraulic system being controlled.Due to large energy losses in hydraulic actuators, havingthree hydraulic actuators working at once is most likelynot the most efficient solution. Another problem is thatthe solution does not take into account the geometry of themanipulator, i.e. the trajectories of individual links werechosen somewhat arbitrarily. For example, having the firstlink rotating at high velocity requires high energy, whereasextending the boom section at high velocity takes lessenergy due to the difference in the inertias. An interestingtask to consider is optimizing the trajectories in terms oftime, energy loss, oil consumption, or other costs.

6. DIRECT HYDRAULIC CONTROL

By measuring the pressures in the hydraulic cylinderchambers it is possible to counteract oscillations muchmore directly than via measurements of link positions. Inthis section, results on damping of oscillations in the first

2 2.25 2.5 2.75 3 3.25 3.5 3.75 4−1

−0.75

−0.5

−0.25

0

0.25

0.5autonomous motion − circle

x−position [m]

z−po

sitio

n [m

]

Fig. 6. Desired reference trajectory vs. performed boom-tip trajectory.

Hydraulic

Actuator

Mechanical

Device

Position

ReferencePosition

Position

Controller

Torque

Controller

Current Torque

Fig. 7. Two stages cascade control.

link of the crane are shown. The analysis is concentratedon this link since it showed to be the main cause of suchan oscillatory behavior. Besides, the absence of pressuresensor devices does not allow us to reproduce the referencetracking system described above. However, such systemsare currently under preparation.

The idea is to have a cascade control scheme to control thehydraulic force and angular position simultaneously, whichwill result in reduced oscillations while asymptoticallystabilize the desired force and motion.

Fig. 7 shows a block diagram describing the scheme of thecascade control. As it can be seen, the control problemis split into two stages. The outer loop controller calcu-lates the reference piston forces Fref needed to drive themanipulator along the predefined joint space θref . Theinner loop controller takes this force reference Fref andcomputes the servo-valve input current u needed to makethe true piston forces F asymptotically track Fref . SinceF asymptotically tracks Fref , which is itself an asymptoti-cally stabilizing control for the manipulator motion aroundθref , the overall cascade system is asymptotically stablearound the trajectory reference θref .

6.1 Hydraulic force model

Since modeling the hydraulic force from first principles isfar from trivial, due to nonlinearities and the very largenumber of parameters, in this work it is shown how to usesystem identification methods to find a transfer functionrelating the input current (to the four-way valve) withthe hydraulic force produced by the cylinders. However,the model structure and order are chosen based on first-principles analysis, which are described below.

Fig. 8. Double acting piston-type actuating cylinder.

Disregarding internal disturbances, nonlinearities causedby friction, etc, the applied force produced by the linearhydraulic actuator shown in Fig. 8 is given by [11]

F = AApA − ABpB (2)

where AA and AB are the areas at each respective cham-ber, pA and pB are the pressures measured at each cham-ber. The relations governing the dynamics of the pressuresare [11]:

pA = βVA

[−CempA − AAxp + qA],

pB = βVB

[−CempB + ABxp − qB ],(3)

where VA and VB are the chambers volumes, Cem is thecylinder internal leakage, xp is the pistons displacement,qA and qB are the input and out flow to and from thecylinder chambers. The linearized version of the hydraulicflow for qA and qB is [11], [12]:

qA = 2Kqxs − 2Kc(pA − ps/2)qB = 2Kqxs + 2Kc(pB − ps/2)

(4)

where Kq and Kc are known as the valves coefficients, ps

is the supplied pump pressure and xs denotes the motionof the four-way valve spool displacement.

The relation between the four-way valve spool position xs

and the input current u can be written as [12]

xs(s) =ω2

n

s2 + 2ξωn + ω2n

u(s) (5)

where ξ and ωn represent the damping ratio and naturalfrequency characteristics of the servo valve.

By combining equations (2) to (5) and by considering thatthe system parameters represent some numerical values, alinear model is obtained, in which parameters have beencollected and substituted to simplify notation,

Fcyl(s) =AA·cω2

n(s+f)+AB ·hω2

n(s+a)(s+a)(s+f)(s2+2ξωn+ω2

n) · u(s)

+ s(−AA·b(s+f)−AB ·g(s+a))(s+a)(s+f) xp(s)

+AAe(s+f)−ABm(s+a)(s+a)(s+f) .

(6)

Based on this analysis, it is shown that the hydraulictorque can be modeled by a fourth-order linear timeinvariant system.

6.2 Hydraulic system identification

Closed-loop identification was used to find the hydraulicdynamics. Different trajectories were designed and tracked

0 5 10 15 20 25 30 35 40−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3

Des

ired

θ 2(t),

(re

fere

nce)

time(sec)

1

2

5

4

3

6 7

8

Fig. 9. Desired trajectory for the first link position.

0 0.5 1 1.5 2 2.5 3 3.5 4−4000

−2000

0

2000

4000

6000

8000

Time

y1

0 0.5 1 1.5 2 2.5 3 3.5 4−0.05

0

0.05

0.1

0.15u1

Time

Fig. 10. Force response to the desired trajectory. The graphon top shows Force and bottom shows input current.

using the position control presented in Section 5. Theresulting valve current resembles a series of step responses.In Fig. 9 one example of such trajectory is presented,and in Fig. 10 the response of the force and the inputcurrent for the first step (denoted by 1 in Fig. 9), fromthis trajectory, is shown.

An analysis of the Fourier transform and spectral analysis(see Fig. 11 and 12) reveals the second order characteristicsof the valve dynamics (5).

Fourth order systems were identified using several methodsincluding the prediction error method (PEM) and theoutput error (OE) method. The best validation resultswere obtained with PEM. The transfer function found was:

Fpem(s) =b3s

3 + b2s2 + b1s + b0

a4s4 + a3s3 + a2s2 + a1s + a0(7)

where b3 = 2.123×105, b2 = 4.253×107, b1 = 4.579×108,b0 = 6.197 × 108, a4 = 1, a3 = 15.68, a2 = 647.2,a1 = 8636, a0 = 2.353 × 104.

Properties such as DC gain, damping and natural fre-quency in terms of bode diagram is depicted in Fig. 13for different methods.

101

102

104

105

106

Am

plitu

de

From u1 to y1

101

102

−300

−200

−100

0

100

Pha

se (

degr

ees)

Frequency (rad/s)

g_spa75g_spa100g_spa125

g_spa75g_spa100g_spa125

Fig. 11. Spectral density analysis of recorded data usingdifferent hamming windows.

0 5 10 15 20 25 30 35 40 45 500

1

2

3

4

5x 10

6

FF

T o

f For

ce(t

), (

outp

ut)

0 50 100 150 200 2500

20

40

60

80

100

FF

T o

f u(t

), (

curr

ent)

frequency [rad/sec]

Fig. 12. Fast Fourier transform of the recorded data.

10−1

100

101

102

103

102

104

106

Ampl

itude

Bode plots for identified models

10−1

100

101

102

103

−500

−400

−300

−200

−100

0

100

Phas

e (d

egre

es)

Frequency (rad/)

m4coe340oe340

2

m2cg_spa125

m4coe340oe340

2

m2cg_spa125

Fig. 13. Bode Diagram from different identification meth-ods super-imposed to the spectra analysis of the initialdata.

6.3 Control design for the hydraulic force

The linearized models found by identification are an ap-proximation of the real system dynamics. Many of the pa-rameters will vary depending on temperature, componentage, and other factors, and it is important to design acontroller which is robust to such uncertainties [10].

1 1.5 2 2.5 3 3.5 4 4.5 5 5.5−20.000

0

20.000

40.000

60.000

time [s]

Cylin

de

r F

orce

[N

]

Force Control for the Hydraulic Actuators

Feedback ControlOpen Loop

Fig. 14. Force control experimental result.

The H∞ loop-shaping design method is an effective way todesign a robust controllers to meet our requirements. Theapproach is based in choosing a particular designed closed-loop response, and then finding an optimal controllerwhich minimizes the H∞ norms of weighted performancemeasures [13, 14].

A simple target loop shape is a first order low pass filterof the form Gcl(s) = 1/(s/Wc + 1), where Wc representsthe cutoff frequency. The resulting controller, for a targetloop with Wc = 3, has the transfer function:

C(s) =c4s

4 + c3s3 + c2s

2 + c1s + c0

l5s5 + l4s4 + l3s3 + l2s2 + l1s + l0(8)

where c4 = 0.02579, c3 = 0.4043, c2 = 16.69, c1 = 222.7,c0 = 606.7, l5 = 1, l4 = 4308, l3 = 8.373×105, l2 = 1.133×107, l1 = 3.854 × 107, l0 = 3.594 × 107.

In Fig. 14 some experimental results are shown. In thisspecific result the force of the first link was recorded whileperforming some motion (dotted line) under only positioncontrol, while the second plot shows the pressure under acascade control configuration. It can be said that very largeoscillations were damped effectively using this method.

7. CONCLUSIONS AND FUTURE WORK

The primary target of this project is to increase produc-tion efficiency of forestry machines by means of intelligentcontrol techniques. Essential to any autonomous or semi-autonomous forestry crane is good low-level system identi-fication and control. In this paper we have presented somepromising preliminary results along these lines. The mainproblems overcome were related to friction identificationand compensation, and damping of oscillations within thehydraulic system.

There is much that can be done to continue this project. Atpresent the cascade-control structure, which showed verypromising results in damping of oscillations, has only beenimplemented on the first link of the crane. Extending thisto all links is a natural evolution of the current work, and itis expected that this will allow much smoother trajectorytracking to be achieved.

Other related work includes using the proposed low-level in concert with higher-level control schemes, suchas semi-autonomous teleoperation systems (see, e.g., [3]),and fully-autonomous systems based on optimized motionplanning.

REFERENCES

[1] U. Hallonborg, “Forarlosa skogsmaskiner kan blilonsamma (Unmanned forestry machines can be com-petitive)” (in Swedish), Skogforsk Results, No 9, 2003.

[2] M.Brander, D. Eriksson, B. Lofgren, “Automation avkranarbetet kan oka prestationen (Automation ofknuckleboom work can increase productivity)”, Skog-forsk Results, No 8, 2004.

[3] S. Westerberg, I. R. Manchester, U. Mettin, P. LaHera, A. Shiriaev, “Virtual Environment Teleopera-tion of a Hydraulic Forestry Crane”, accepted for the2008 IEEE International Conference on Robotics andAutomation, Pasadena, CA.

[4] P. D. Lawrence, , F. Sassani, B. Sauder, N. Sepehri,U. Wallersteiner, J. Wilson, “Computer-assisted con-trol of excavator-based machines”, International Off-Highway and Powerplant Congress and Exposition,Milwaukee, Wisconsin, 1993.

[5] J. Mattila, T. Virvalo (a), “Computed force controlof hydraulic manipulators”, 5th Scandinavian Inter.Conf. On Fluid Power pp. 139154, 1997.

[6] M. Linjama, The Modelling and Actuator Space Con-trol of Flexibile Hydraulic Cranes, PhD thesis, Tam-pere University of Technology, Tampere, Finland,1998.

[7] M. E. Munzer, Resolved Motion Control of MobileHydraulic Cranes, PhD Thesis, Aalborg University,Denmark, 2002.

[8] M. E. Munzer, P. Pedersen, “Real time simulationmodel of flexible mobile crane for machine-operatorinteraction testing” Proc. 2nd International Work-shop on Computer Software for Design, Analysis, andControl of Fluid Power Systems, 2001.

[9] H. Olsson, K. J. Astrom, C. Canudas de Wit,M.Gafvert & P. Lischinsky, “Friction Modelsand Friction Compensation”. European Journal ofControl, Dec. 1998, no.4, pp. 176-195.

[10] F. Bu and Bin Yao, “Adaptive robust precision mo-tion control of single-rod hydraulic actuators withtime varying unknown inertia: a case study”, In Proc.of American Control Conference, 4129-4133, 2000.

[11] Noah D.Manring. Hydraulic Control Systems. JohnWiley & Sons, New York, 2005.

[12] J.A.F. Ferreira, Modelacao de Sistemas Hidraulicospara Simulacao com Hardware-in-the-loop. PhD The-sis (in Spanish), University of Aveiro, Portugal, 2003.

[13] Kemin Zhou, John C. Doyle, Essentials Of RobustControl, Prentice Hall, 1997.

[14] I. R. Petersen, V. A. Ugrinovski, and A. V. Savkin,Robust Control Design using H∞ Methods, Springer-Verlag, London, 2000.

Virtual Environment Teleoperation of a Hydraulic Forestry Crane

Simon Westerberg∗, Ian R. Manchester∗, Uwe Mettin∗, Pedro La Hera∗, Anton Shiriaev∗,†

Abstract— A teleoperation system has been developed for ahydraulic crane, of the type used on aforwarder vehicle, whichtravels off-road and collects logs cut by aharvester. The systemdeveloped consists of a 3D virtual environment, which allowsthe user to input a desired position for the crane tip usingeither the mouse or a joystick. The desired position is thentransmitted (via UDP/IP) to a local control system. The crane isa redundant manipulator, so movements of the individual linksare calculated using a pseudoinverse method, and controlledusing PIDs with friction compensation. Encoder data from thecrane links are continuously sent back to the user side, and thecrane’s movement is visualized in the virtual environment.Thesystem has been tested on a real forwarder crane, experimentalresults and a video of the system’s performance are provided.

I. I NTRODUCTION

The Swedish forestry industry has a long-term goal of de-veloping autonomous and semi-autonomous forestry vehicles[1], [2].

There are two main types of off-road vehicle used in theforestry industry: aharvester, which fells and delimbs thetrees, and cuts the trunk into logs of a predetermined size,and aforwarder, which collects the logs in a tray, and carriesthem to the nearest road for collection. A forwarder is shownin Figure 1. Control of the forwarder can be divided into twodistinct tasks: navigation of the vehicle itself, and operationof the on-board crane. It is the latter task which we considerin this paper.

This paper describes implementation and experimentaltesting of a virtual-environment-based remote control andvisualization system for such a crane. The system presentedhere is to be considered a partial solution. At present, thevirtual environment is constructed based on information fromthe sensors on board the crane, and does not include anydynamic sensing of objects within the environment. Thelatter is currently under development.

A. Virtual environment–assisted teleoperation

Virtual environments systems have been successfully usedfor teleoperational tasks. Gravezet al. [3] use a virtualmodel of a hydraulic arm and its surroundings to performdifferent tasks in a radioactive environment. Their systemallows visual feedback as well as high-level motion control.There are several other reports of successful implementationof virtual-environment-based teleoperation systems for other

This work has been supported by the Centre for Intelligent Off-Road Ve-hicles at Umea University, Komatsu Forest AB, and the KempeFoundation.

* Department of Applied Physics and Electronics, Umea University, SE-901 87, Sweden.

† Department of Engineering Cybernetics, Norwegian University ofScience and Technology, N-7491 Trondheim, Norway.

Fig. 1. A forwarder: the Komatsu 860.1.

classes of systems in the literature, see, e.g. [4], [5], [6]andreferences therein.

The use of virtual environments in teleoperation has sev-eral advantages compared to teleoperation with streamingvideo as visual feedback [7]. One obvious advantage is thepossibility to change the view of the virtual environment.When video is used for feedback, the physical cameras needto be attached to the machine and placed where they do notrisk to be damaged by collisions. The virtual environment,however, can be fused from many different sensors at differ-ent locations, and the virtual cameras can be placed in anylocation and moved around freely. This gives the user a goodoverview and also the possibility to adapt the view to the taskthat is currently performed. E.g., if an obstacle is present, theoperator can view the scene from another direction so thatthe line-of-sight is not blocked by the obstacle.

Additionally, the communication of streaming video re-quires large bandwidth in order to have reasonable imagequality. This is an important issue in the forest, where In-ternet connections have a very limited capacity. With virtualenvironments, the information flow between the machine andthe operator can be restricted to the most relevant featuresof the environment, such as the position and size of alog, whereas less important details in e.g. textures can bediscarded.

Another advantage with virtual environments is the pos-sibility to easily add visual operator assisting features.For

Fig. 2. The system architecture

instance, irrelevant details about the environment can be hid-den, while more important information can be emphasized,in order to let the operator focus on important information.Furthermore, information that is not visual in the physicalworld, like optimal crane-arm trajectories, can be added. Bychanging the modality of the information, non-visual objectproperties can be made visible in the virtual environment.E.g., color can be used to represent the temperature of anobject.

This makes virtual environment useful not only for tele-operation. Even in normal operation, a virtual environent cangive the operator information that is hard or impossible tosee with his own eyes.

A further goal is to make the control of the crane moreintuitive. In current forest machines, the driver must directlycontrol six hydraulic valves using two 3-degree-of-freedomjoysticks. Each joystick motion corresponds to one cranefunction, e.g. increasing the angle between two links, orextending the telescopic arm, etc. Learning to control thecrane tip efficiently takes years of training [1]. A moreintuitive way to control the crane is to directly specify theposition of the crane tip, and generate the required hydraulicsinputs by computer control.

When certain tasks are automated and thus performedwithout human intervention, a virtual environment is usefulfor monitoring the machine’s actions. Furthermore, even ifacrane is completely controlled by a driver, a virtual environ-ment allows a forestry company to monitor the operation ofa large fleet of vehicles in an intuitive manner.

II. SYSTEM ARCHITECTURE

Figure 2 shows the architecture of the teleoperation sys-tem. The system consists of two subsystems, connectedthrough an IP network, either a LAN or the Internet.

One subsystem is at the location of the crane, with physicalconnections to sensors and actuators. Besides the sensor-equipped crane and the real-time device, this subsystemcontains an ordinary PC. The PC is responsible for the com-munication between the real-time system and the network.

The second subsystem is at the location of the operator.It provides the operator with visual feedback, as well aswith an interface for controlling the crane. For visualization

Fig. 3. The crane laboratory at Umea University.

and user interface, the CraneVE software was developed.The software is written in C++ and uses a scene graph asthe virtual environment data structure, using the OpenScene-Graph library. This subsystem consists of a PC that runsthe CraneVE software as well as the software that readsinput from the user. The operator uses a joystick and/or amouse to enter the input to the system. The output consistsof visualization on a monitor.

The crane PC and the operator PC are connected throughan IP network. Data transmissions are sent using UDP, whichhas a low overhead and reduces the bandwidth use. In onedirection, the transmitted data consists of sensor data, i.e.link angles and telescope extension. In the other direction, itconsists of 3D positions that are used to control the crane.Since only a few bytes are used to represent the transmittedinformation, the bandwidth requirement is low, even if thesensor data are updated frequently.

III. C RANE CONTROL HARDWARE

The following experimental equipment has been installedat a laboratory at Umea University:

• CRANAB 370RCR hydraulic forwarder crane,• three joint-position encoders, 4000 pulses per turn,

the telescope extension is measured by an encoderconnected to a retracting wire,

• dSPACE MicroAutoBox real-time protyping hardware,with on-board PowerPC CPU,

• RapidPro power amplification unit and signal condition-ers.

The 370RCR crane is somewhat smaller than most cranesused on real forwarders, but was chosen due to spacerestrictions in the laboratory. It’s operational principle anddynamics are, however, very similar. The experimental setupis shown in Figure 3.

IV. U SER INTERFACE

The user interface is shown in Figure 4. The operator ispresented with a view of the virtual model of the crane, aswell as of the surrounding environment.

Fig. 4. The user interface for the crane teleoperation system.

On the left side is the view of the virtual environment. Theviewpoint can be moved around freely by the user using themouse: the operator can move and rotate the camera, as wellas zoom in and out. This allows the user to zoom out foran overview, or focus on details of the environment that areimportant for a certain task.

The other two view panes have positions and directionsthat are fixed relative to the crane. That is, for any config-uration (rotation and translation) of the crane, the camerasare always directed towards the crane. The top right viewpane shows the crane from the side, while the bottom rightview pane shows the crane from above. This allows theoperator to always have a detailed view over the crane andthe environment near the crane.

Virtual environments are suited for the introduction ofoperator assisting features. As one example of this, visual-ization of the crane workspace has been implemented. Thiscan be seen in Figure 5.

A. Target Specification

The user can specify the target position of the crane-tipin a number of different ways. One way is to use the mouseto click on the top right camera view. This will result ina position in the vertical plane. The second method is touse an ordinary joystick to move a pointer in the virtualenvironment. A 2-axis joystick is used and two buttons areused to simulate a third axis for full 3D motion. When theuser presses a button, the current position of the pointeris sent to the control system. The crane motion that isrequired to move the crane tip from the current crane positionto the target point is then calculated. The motion can becalculated to be optimal with regard to e.g. speed or energyefficiency. In the following section we describe a simplemotion-planning algorithm which can be calculated on-line.If several subsequent target points are specified, the targetpoints become waypoints on a target trajectory. The cranetip will then visit each waypoint in order.

The next step is to transfer more responsibility from the

Fig. 5. High-level control with collision avoidance. The yellow path showsthe trajectory generated by the collision avoidance algorithm in order toavoid the obstacle (the green box) while moving towards the tray.

operator to the control system. This means that certain tasksor processes are automated and performed autonomously bythe system. As an example of an activity where the computerhas increased responsibility, a simple collision-avoidancealgorithm has been implemented in the CraneVE software.When collision avoidance is enabled, the system performscollision detection between the user-specified trajectoryandthe objects in the scene. If the user specified trajectory isblocked by an object, a new trajectory is calculated suchthat the crane tip passes above the object and avoids thecollision. The collision avoidance algorithm is explainedinFigure 6.

V. CRANE-SIDE CONTROL

The configuration of the crane is represented by a three-dimensional vector,q = [θ2, θ3, d]′. From basic trigonometryone can derive the forward kinematics equations, giving theCartesian coordinates from a known link configuration:

c := [x, y]′ = f(q). (1)

However, the low-level control objective is to reach a particu-lar point in two-dimensional space, i.e. a particular Cartesiancoordinatec? = [x?, y?]

′. Since there are three degrees offreedom in the crane, this is an underdetermined problem,and hence has infinitely many solutions. We now describethe method we use to choose a correspondingq? suchthat f(q?) = c?, it is conceptually similar to the methodspresented in [8], [9], [10].

Differentiating the above equation, we get:

c = F (q)q, (2)

whereF (q) is the Jacobian off(q). We wish to calculatea movement inq that would correct the current positioningerror c? − c.

Again, generatingq from c is an underdetermined prob-lem, and has infinitely many solutions. We choose to find

Fig. 6. The collision avoidance algorithm. (a) Assume thatp is the end pointof the current path. Letq be a new target point defined by the operator. (b)If q collides with an object, find a safe pointq′ above the bounding sphereof the object. (c) If the pathp → q′ collides with an object, find a safepoint c′ above the center of the object. Letm and n be the middle pointof each sub-pathp → c′ andc′ → q′ respectively. (d) Find safe pointsm′

and n′ abovem and n such that they are outside of the bounding sphereof the colliding object. The final path will bep → m′ → c′ → n′ → q′.

Fig. 7. Basic geometry of the crane.

the solution which minimizes a configuration-dependent costfunction of the form:

J = q′P (q, c)−1q, (3)

whereP (q, c) is a positive definite matrix for allq and c.This choice of cost function is quite simplistic, but it has

the advantage that the resulting solution can be calculatedanalytically:

q = P (q, c)F (q)′[F (q)P (q, c)F (q)′]−1c. (4)

The matrixF (q)P (q, c)F (q)′ is three-by-three, and its in-verse can be expressed analytically in terms of its elements.The formulas are too lengthy to present here, but they arenot difficult for real-time computation, in comparison to aniterative scheme that another choice of cost function wouldrequire.

In our system,P (q, c) is chosen to be of the followingform:

P (q, c) =

w1p1(q, c) 0 00 w2p2(q, c) 00 0 w3p3(q, c)

. (5)

The valuesw1, w2, w3 represent the desired contributionfrom each link to the overall movement. If one of them iszero, the corresponding link is stationary and the other twolinks are used to generate the motion.

The functionsp1(q, c), p2(q, c), p3(q, c) are there for “pro-tection”: if one of the links is approaching the limit of itsworking range, its contribution should be smoothly reducedtowards zero, and the other links should take over. We nowdescribe how they are calculated.

For eachi = 1, 2, 3, let qmaxi and qmin

i be the maximumand minimum allowed values ofqi, let δi be a constantsmaller than(qmax

i − qmini )/2, and letφi(q) be a smooth

function such that:

• φi(qmini ) = 0,

• φi(qmaxi ) = 0,

• φi(s) = 1 for all s ∈ [qmini + δi, q

maxi − δi]

Thenpi(q, c) is calculated like so:

Algorithm 1:

1) Setp1(q, c) = p2(q, c) = p3(q, c) = 1,2) calculateq0

? from Equations (4) and (5),3) for eachi = 1, 2, 3, if

• qmaxi − qi < δi and q0

?,i > 0,or• qi − qmin

i < δi and q0?,i < 0

thenpi(q, c) = φi(qi).

�

We have then the following closed-loop control algorithm,which generates control signalsu1, u2, u3 from a desiredCartesian set-pointc? and the measurementsq1, q2, q3:

−1 0 1 2 3 4 5 6 7 80.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Time (s)

Firs

t lin

k an

gle

(rad

)

Fig. 8. Step response for the first link,θ2.

Algorithm 2:

q?(0) = q

q? = kP (q, c)F (q)′[F (q)P (q, c)F (q)′]−1(c? − f(q?)),

u1 = C1(s)[q1 − q?,1],

u2 = C2(s)[q2 − q?,2],

u3 = C3(s)[q3 − q?,3].

(6)

wherek is a large positive gain, andC1(s), C2(s), C3(s) arePID controllers tuned for each link of the crane, andP (q, c)is calculated at each sample time from Algorithm 1.�

Each control signal is further modified to compensatefor the large Coulomb friction present in the hydraulicsactuators, see [11] for details, see also [12] for similar work.

The vector q? quickly converges to a feasible target,satisfyingf(q?) = c?. Following this, the individual link po-sitionsq1, q2, q3 are driven more slowly to the correspondingvalues fromq?.

VI. EXPERIMENTAL RESULTS

A. Individual Link Positioning

In Figures 8, 9, and 10 we see the step responses ofthe individual links, as controlled by the PID controllersC1(s), C2(s), C3(s) along with the friction compensators.

It is clear that the first link is by far the slowest, which isreasonable since it must drive by far the greatest mass. Thesecond link and the telescope converge at similar rates, butthe telescope is much smoother and easier to control, sinceit has the least mass to drive.

We chose the control contributions in Algorithm 2 suchthat the links with the best control are used most. Thefollowing values were found to be reasonable in experiments:w1 = 0.5, w2 = 1, w3 = 1.5.

Overall, we were able to make all links converge tothe desired positions with zero steady-state errors, sufficientspeeds, and without oscillations.

−1 −0.5 0 0.5 1 1.5 2 2.5 3−1.6

−1.5

−1.4

−1.3

−1.2

−1.1

−1

−0.9

Time (s)

Sec

ond

link

angl

e (r

ad)

Fig. 9. Step response for the second link,θ3.

−1 −0.5 0 0.5 1 1.5 2 2.5 30

0.2

0.4

0.6

0.8

1

1.2

1.4

Time (s)

Tel

esco

pe p

ositi

on (

m)

Fig. 10. Step response for the telescope,d.

B. Teleoperation Experiment

Accompanying this paper is a video of the teleoperationsystem in action (a higher resolution version is also availableonline [13]). The user clicks the mouse at a sequence ofpoints in the virtual environment, and the crane can beobserved moving to those points. Figure 11 shows thex andy setpoints and crane trajectories from the same experimentas featured in the video.

It can be seen that in general the system converges quickly(in two to four seconds) to the desired setpoint.

There is an interesting effect visible at around 17 seconds,and again at 30 seconds. When the desired motion is largein one direction, but small or zero in the other, we still canobserve a noticable deviation in the latter dimension. Thisisbecause, with a largek in Algorithm 2, q? converges veryquickly to its final value satisfyingf(q?) = c?, and theneach link is driven to its desired value at a slower rate. Inparticular, the first link is much slower than the other two.This means that the end-effector will not in general followa straight line path between two points in Cartesian space.

VII. C ONCLUSIONS ANDFUTURE WORK

The virtual environment-based teleoperation system re-ported in this article can be considered proof of concept.The system is also useful as a tool for further research anddevelopment in the area of remote crane control.

15 20 25 30 35 40 45 502

2.5

3

3.5V

ertic

al p

ositi

on (

m)

15 20 25 30 35 40 45 502

2.5

3

3.5

4

4.5

Time (s)

Hor

izon

tal p

ositi

on (

m)

TargetTrue position

Fig. 11. Horizontal and vertical positioning during the teleoperationexperiment.

However, the system is far from ready to be used in realforest machines. A virtual environment-assisted teleoperationsystem depends on the virtual representation being accurate,and the main problem right now is the lack of knowledgeabout obstacles in the environment. In order to correct this,we need a reliable method to detect and classify differentobjects near the crane. The classification problem is quitedifficult: the system must be able to distinguish betweenrocks, trees, logs, branches and mud. It also needs to reliablydetect any human inside the safety area.

A number of different technologies are under investigationto address this problem. Different sensor systems like stereocameras, laser scanners and structured light systems [14] canbe used to extract information from the environment.

Another direction of our current work is to implement andtest a similar system on a real forwarder crane, provided byKomatsu AB. Due to reasons of confidentiality, we are notable to publish the results of these experiments, but a videoshowing an early test is available online [15].

REFERENCES

[1] U. Hallonborg, “Forarlosa skogsmaskiner kan bli lonsamma (Un-manned forestry machines can be competitive)” (in Swedish), Skog-forsk Results, No 9, 2003.

[2] M. Brander, D. Eriksson, B. Lofgren, “Automation of knuckleboomwork can increase productivity”,Skogforsk Results, No 4, 2004.

[3] P. Gravez, C. Leroux, M. Irving, L. Galbiati, A. Raneda, M. Siuko, D.Maisonnier, and J. D. Palmer, “Model-based remote handlingwith theMAESTRO hydraulic manipulator”,Fusion Engineering and Design,vol. 69, no. 1, pp. 147–152, 2003.

[4] L. Fluckiger, L. Piguet, and C. Baur, “Generic robotic kinematicgenerator for virtual environment interfaces”, inSPIE Telemanipulatorand Telepresence Technologies, Vol. 2901, pp. 186–195, 1996.

[5] R. Safaric, R. M. Parkin, C. A. Czarnecki, and D. W. Calkin, “Virtualenvironment for telerobotics”,Integrated Computer-Aided Engineer-ing, vol. 8, no. 2, pp. 95–104, 2001.

[6] L. A. Nguyen, M. Bualat, L. J. Edwards, L. Flckiger, C. Neveu, K.Schwehr, M. D. Wagner, and E Zbinden, “Virtual reality interfaces forvisualization and control of remote vehicles”,Autonomous Robots, vol.11, no. 1, pp. 59–68, 2001.

[7] A. Kheddar, R. Chellali, and P. Coiffet, “Virtual environment-assistedteleoperation” inHandbook of Virtual Environments: Design, Imple-mentation and Applications, 2001, ch. 48, pp. 959-998.

[8] A. A. Mohamed and C Chevallereau, “Resolution of Robot Redun-dancy in the Cartesian Space by Criteria Optimization”, inProceedingsof the IEEE International Conference on Robotics and Automation,Atlanta, GA, 1993, pp. 646-651.

[9] L. Beiner, “Minimum-Force Redundancy Control of HydraulicCranes”,Mechatronics, vol. 7, no. 6, pp. 537–547, 1997.

[10] L. Beiner and J. Mattila, “An improved pseudoinverse solution forreduntant hydraulic manipulators”,Robotica, vol. 17, pp. 173–179,1999.

[11] P. La Hera, U. Mettin, I. R. Manchester, A. Shiriaev, “Identificationand Control of a Hydraulic Forestry Crane”,submitted to the 2008IFAC World Congress, available on request.

[12] M. E. Munzer, “Resolved Motion Control of Mobile HydraulicCranes”, PhD Thesis, Aalborg University, Denmark, 2002.

[13] Video of crane virtual environment, available online atwww.tfe.umu.se/forskning/Control Systems/Movies/virtual environment exp.AV I

[14] S Lee, J Choi, D Kim, J Na, S Oh, “Signal Separation CodingforRobust Depth Imaging Based on Structured Light”, inProceedingsof the IEEE International Conference on Robotics and Automation,Barcelona, 2005.

[15] Video of experiment on forwarder, available online atwww.tfe.umu.se/forskning/Control Systems/Movies/forwarder crane exp.AV I