Sie sind auf Seite 1von 26

A Mechanical Model for Robotic Joints with Harmonic Drives

Wolfgang Sey erth, Jorge Angeles TR-CIM-94-13 February 8, 1995

Department of Mechanical Engineering & Centre for Intelligent Machines McGill University Montreal, Quebec, Canada

Postal Address: 817 Sherbrooke Street West, Montreal, Quebec, Canada H3A 2K6 Telephone: (514) 398-6315 Fax: (514) 398-7348 Electronic Mail: angeles@cim.mcgill.ca

A Mechanical Model for Robotic Joints with Harmonic Drives


Wolfgang Sey erth, Jorge Angeles

Abstract
Harmonic drive gears, of extensive use in robotics, exhibit a non-smooth torque transmission characteristic that signi cantly in uences the dynamic behaviour of the driving and driven elements. For the design of robust motion- and force-control schemes, it is therefore necessary to model accurately the indirect drives involved. In this report, we analyze nonlinear harmonic drive properties, that have been observed in experiments, and propose a mechanical model for the joint transmission unit. The main sources of nonlinearity, that have also been taken into account in the modelling, are identi ed as compliance, dynamic friction, static friction (startup torque) and hysteresis, the last two items also depending on the load history of the harmonic drive due to frictional memory. Based on a series of quasistatic and stationary dynamic experiments, we then describe a scheme to identify all necessary parameters. Finally, the quality of the model and parameters is veri ed experimentally with respect to its dynamic behaviour by driving the actual system and the model with the same input and comparing the output signals. Numerical and experimental results are given for two harmonic drives and illustrate that the nonlinear system dynamics can be accurately predicted.

Acknowledgements
The research work reported here was completed with the support of the Natural Sciences and Engineering Research Council of Canada and a Feodor Lynen Research Fellowship, granted by the Alexander von Humboldt Foundation to the rst author.

Une Modele Mecanique pour des Articulations Robotiques a Harmonic Drives


Wolfgang Sey erth, Jorge Angeles

Resume
Les reducteurs harmoniques, d'un emploi tres rependu en robotique, possedent une caracteristique irreguliere de transmission des moments qui in uence considerablement le comportement des actuateurs et des elements commandes. Pour la conception d'un algorithme de commande robuste en force et en position, il est necessaire de modeliser tres precisement la commande. Dans ce rapport, nous analysons les proprietes non-lineaires des moteurs harmoniques, a partir de donnes experimentales, tout en proposant un modele mecanique de l'articulation de transmission. Les principales sources de non-linearites, prises en compte dans la modelisation, sont identi ees comme la exibilite, le frottement dynamique, le frottement statique (moment de demarage) et l'hysteresis. Ces deux derniers elements dependent egalement de l'historique de charge due a l'e et de memoire de frottement. A partir d'une serie d'experiences quasi-statiques et dynamiques stationnaires, nous decrivons un moyen d'identi er tous les parametres necessaires a l'etablissement du modele. Finalement la qualite de la modelisation et de ses parametres est veri ee experimentalement, en respectant le comportement dynamique, par la comparaison des signaux de sorties du systeme reel et du modele de commande obtenus a partir des m^mes e excitations. Des resultats numeriques et experimentaux sont donnes pour deux types de reducteurs harmoniques illustrant la possibilite de predire e cacement ce type de systemes dynamiques non-lineaires.

Remerciements
Cette recherche a ete nanciee par le Conseil de Recherche en Sciences Naturelles et en Genie et une bourse de recherche Feodor Lynen, accordee par la fondation Alexander von Humboldt au premier auteur.

1 Introduction 2 Harmonic Drive Transmission Properties 3 Mathematical Model


3.1 Model Equations : : : : : : : 3.2 Parameter Identi cation : : : 3.2.1 Harmonic Drive RH-25 3.2.2 Harmonic Drive RH-32 3.3 Model Evaluation : : : : : : :

Contents

2.1 Description of Testbed : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 3 2.2 Experimental Analysis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 4

1 3 7

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

: : : : :

7 9 9 12 15

4 Model Veri cation and Results 5 Conclusions References

4.1 Results for the Harmonic Drive RH-25 : : : : : : : : : : : : : : : : : : : : 17 4.2 Results for the Harmonic Drive RH-32 : : : : : : : : : : : : : : : : : : : : 19

17 21 22

1 Introduction
Harmonic drive (HD) transmissions have gained wide acceptance in industrial applications because of their unique gearing con guration that o ers compactness, light weight, high reduction ratios and virtually zero backlash. On the other hand, HD transmissions bring about nonlinear properties that play an important role in the overall dynamic response of the plant. The design of high-performance motion controls for robotic applications requires an accurate knowledge of the torque characteristic of the drives involved. This investigation is part of a project involving motion and force control of a seven-axes manipulator. The manipulator, called REDIESTRO, has been designed at the Robotic Mechanical Systems Laboratory of the McGill Centre for Intelligent Machines, and is equipped with harmonic drives at all seven joints 1]. While the natural compliance of the joints is advantageous for force- and torque-control applications, it brings about side e ects that must be taken into account for an acceptable performance. Previously published models have been set forth to characterize the nonlinear kinematic and kinetic behavior of harmonic drives. High-resolution quasistatic experiments have been undertaken to study kinematic inaccuracies in the output rotation that are mainly in uenced by manufacturing and assembly imperfections, such as tooth-placement errors, out-of-roundness and misalignments. In 2] it is shown that the position error displays a primary component twice every wave-generator revolution along with additional smaller components at subsequent harmonics. According to manufacturing data, the peak-to-peak amplitude of the kinematic error rarely exceeds 0.033 degrees of output rotation, which has also been con rmed 2]. Transmission compliance, resulting from gear-tooth interaction and wave-generator deformation due to high radial forces, is commonly approximated by a nonlinear or piecewise linear sti ness curve 3, 4]. Hysteresis losses caused by meshing gears, however, are mostly ignored, which is commonly based on the assumption that the quasistatic angular position error due to this phenomenon is negligible. Recently developed models also include nonlinear dynamic friction composed of a constant and a velocity-dependent part 5] or a nonlinear sti ness pro le with soft windup and hysteresis that is based on Coulombtype friction 6]. Extensive experiments focusing on kinematic error and frictional e ects have been conducted by Tuttle 7]. Various models, also including Coulomb friction in the gear-tooth area, have been proposed, for which the parameters were painstakingly determined in simulation runs in order to achieve agreement with the actual behavior. The theoretical sti ness pro le thus obtained, however, showed signi cant deviation from experimental sti ness data, pointing out the troublesome nature of system modelling and system identi cation. In this investigation, we focus on the aspects of both model development and parameter identi cation. Speci cally, the mechanical model considered allows for torsional compliance, microscopic friction in the tooth engagement area giving rise to hysteresis e ects, dry (or macroscopic) friction resulting in stick-slip motion of the wave-generator, and velocity-dependent damping. These e ects are studied in two test cases, namely, quasistatic experiments with constrained output rotation and unconstrained stationary dynamic experiments. Experimental observations illustrate that the torque transmission

2 characteristics show frictional memory depending on the operation conditions, and that deviations in the output rotation due to hysteresis e ects can reach up to 0.3 degrees. These memory-dependent properties are captured in the model by parameters that depend nonlinearly on the load history. We then propose a straightforward parameter estimation scheme to complete the model development. Finally, the frequency response is predicted via simulation and compared with the response of the physical system. The comparison of numerical and experimental results shows good agreement and illustrates the quality of our model.

2 Harmonic Drive Transmission Properties


2.1 Description of Testbed
This section provides a brief description of the testing facility on which all experiments have been conducted. The experimental setup for a robotic joint with harmonic drive is shown in Fig. 1. It consists of a DC motor, an elliptical ball bearing assembly (wave generator), an elliptical, nonrigid, external gear ( exspline), and a round, rigid, internal gear (circular spline). The circular spline is xed to ground, while a low-torque, high-speed DC motor drives the wave-generator, that causes the external teeth of the exspline to mesh continuously with the internal teeth of the circular spline along the major axis of the wave generator ellipse. This gear engagement produces a high-torque, low-speed rotation on the exspline or output shaft, repectively, which is directed opposite to the input rotation. For experiments with constrained output motion the output shaft is mounted to ground via a torque sensor whose compliance can be neglected compared to that of the harmonic drive itself. An incremental encoder built within the motor senses the input angular shaft displacement; the input torque can be calculated from the motor current.

torque sensor output shaft support base exspline circular spline wave generator DC motor

encoder Figure 1: Testbed

4 In order to control the motor and monitor sensor output, a data acquisition facility, developed at the Robotic Mechanical Systems Laboratory of the McGill Centre for Intelligent Machines, has been used. The facility consists of a Sparc station 1+ , an SBUS-to-VMEbus adapter, a digital signal processing (DSP) board, D/A and A/D converter boards, ampliers, and an encoder interface. A C-language software environment executed through the host computer speci es the motor current input that is converted at the D/A board and ampli ed by a servo ampli er before being fed to the test station. The output signal from the torque sensor, that is ampli ed by a variable-gain ampli er, and the analog signal from the motor current ampli er are fed through A/D converters to the DSP board, while the motor angle signal is read through an encoder interface. The sensor information, that is collected at a sampling frequency of 300 Hz, is then available on the host computer for further data processing. Information on sensor callibration as well as the handling of the control software can be found in 8].

2.2 Experimental Analysis


The rst step prior to modelling consists in the analysis of experimental data whose signature eventually de nes the structure of the mechanical model. In this investigation, two di erent types of harmonic drives have been tested, RH-25-1507-BE020AL-SP and RH-32-1212-BE036AL-SP, of Harmonic Drive Systems, Inc. 10]. Since both types revealed the same structural input/output behavior, only results obtained from the RH-25 tests will be discussed here. Results for RH-32 are summarized in Subsection 3.2.2.
200

150

100

output torque [Nm]

50

50

100

150

200 1

0.8

0.6

0.4

0.2

0 0.2 torsion [deg]

0.4

0.6

0.8

Figure 2: Measured hysteresis (torsion equals input rotation divided by gear ratio)

5 The rst series of experiments addresses the quasistatic deformation of the transmission unit when the output shaft is blocked. The motor current command is prescribed by a low-frequency sinusoidal function with zero mean, thereby producing a quasistatic motion of the input shaft due to the compliant nature of the harmonic drive. The sti ness pro les obtained from six tests with varying excitation amplitudes are shown in Fig. 2, where the output torque has been ltered with a rst-order, low-band lter in order to eliminate noise in the recorded signal. As we can see, the torsional angle does not return to its initial value after a loading/unloading sequence is applied, which manifests itself in hysteresis losses. The frictional torque that results from small or microscopic relative motion in the teeth contact area, however, is not of the Coulomb type, as there is no jump in the torque history when the direction of input motion changes. On the contrary, the friction losses increase with growing values of maximum applied loads or number of meshing tooth pairs, respectively, this gure showing a dependence on the previous load history instead of actual gearing forces. Furthermore, friction within the meshing teeth may even increase slightly the effective sti ness during small deformation cycles due to the contribution of local contact sti ness in the tooth rubbing area, which can be seen in Fig. 2 by the di erent gradients at vanishing torsion values. The values of the friction losses, that can cause quasistatic deviations of the gear deformation of up to 0:3 , suggest that their in uence on the system dynamics cannot be neglected. The results furthermore indicate that the observed hysteresis behavior cannot be represented by a single sti ness/friction function. Instead, a set of functions must be taken into consideration, that describes the torsion-torque characteristic for di erent loading cycles.
100 80 60 40 20 0 20 40 60 80 100 0.5

torque [Nm]

0.4

0.3

0.2

0.1

0 0.1 torsion [deg]

0.2

0.3

0.4

0.5

Figure 3: Torque vs. torsion (- - output torque, | input torque converted to output shaft)

6 We used the same experiment to observe the torque transmission ine ciency that is manifested by output torques lower than those expected by the commanded input torque, as shown in Fig. 3. These transmission losses result from the startup torque that is required to e ect input rotation causing a stick-slip motion of the input shaft in the experiment. The startup torque, that is characterized by changing input torque while no input rotation is observed, has a typical Coulomb friction signature and also increases with growing load history. The Coulomb friction torque can be traced from Fig. 3 by subtracting the torque sensor output from the motor input torque. Figure 4a shows that the motor friction torque depends mostly both on the direction of torsion and the maximum applied load, while its magnitude is apparently velocity-independent. Damping in harmonic drives manifests itself statically and dynamically. In order to measure the velocity-dependent dynamic friction, additional experiments with unconstrained output shaft and constant angular velocities have been conducted. The damping characteristic in Fig. 4b summarizes the data collected at constant operation velocities related to di erent steady input torques. It shows the no-load starting torque at vanishing angular velocity, as well as the velocity-dependent damping, whose slope decreases with increasing velocities.
0.15

0.2
0.1

o experimental data

0.18 0.16

motor static friction torque [Nm]

0.05

motor friction torque [Nm]


0.4 0.3 0.2 0.1 0 0.1 output torque [Nm] 0.2 0.3 0.4 0.5

0.14 0.12 0.1 0.08 0.06

0.05

0.1

0.04 0.02

0.15 0.5

0 0

50

100

150 200 250 wavegenerator velocity [rad/s]

300

350

400

a) Velocity-independent friction

b) Dynamic friction

Figure 4: Friction Measurement

The goal of modelling a physical plant is to accurately describe essential and observable phenomena while keeping the model representation transparent and as simple as possible, in order to limit computation time. The lumped-parameter model developed here for a harmonic-drive-operated robotic joint with exspline output motion is shown in Fig. 5.

3.1 Model Equations

3 Mathematical Model

J2

output inertia output shaft ( 2) compliance & friction

1 N

J1

gear ratio wave generator ( 1) 6 input inertia motor torque m input damping & startup torque

Figure 5: Mechanical model of gear transmission The input shaft is driven by the motor input torque m that is proportional to the input current im in the non-saturated region of the motor ampli er subsystem, i.e., (1) m = K t im with Kt being the motor torque constant. The input angle 1 is converted to the output side through the constant gear ratio N and the gear deformation is then given by (2) = N1 ? 2 ; where the exspline output rotation 2 has a reference direction opposite to the wavegenerator rotation 1. The transmission compliance acting between input and output shaft is approximated by a cubic function of the torsion angle, i.e., Tc ( ) = c 1 + c 3 3 (3)

8 while the shape of the hysteresis is estimated as a combination of Coulomb friction and a weighted friction function, which is represented by a hyperbolic function 9] as Th( ; _ ) = Tc( ) + Th + Bf ? Th sgn( _ )] tanh( ( ? )) (4) with Th = Th ? Tc

Th ( ; _ )

Th

Th

6 ? 6 ?

Tc( )

Tc

Figure 6: Hysteresis model The Coulomb friction parameter Bf de nes the limits Tc ? Bf Th Tc + Bf of the torque transmission function, whereas the factor determines the slope at the transition from loading to unloading, i.e., _ < 0 ! _ > 0, or vice versa. The hyperbolic function in eq. (4) implies that the friction torque asymptotically approaches a constant friction torque Bf for large arguments ( ? ), so that the parameter determines the contribution of the actual friction, as shown in Fig. 6. The superscript ( ) in eq. (4) relates to quantities at the reversal point from loading to unloading, that can occur at any point ( ; Th ), so that the shape of the torque function (4) also depends on the load history. In contrast to Coulomb-type or macroscopic friction, this hysteresis approach does not contain static friction that blocks the relative motion between input and output rotation. Instead, it aims at friction e ects in the context of microscopic relative motion, where the e ective sti ness is increased at low dynamic loads as a result of the contribution of the tangential sti ness in the local contact area. Besides the transmission characteristic between input and output shaft, there is a power loss during operation due to starting torque and damping in bearings. Since the harmonic

9 drive output operates at relatively small velocities, the dynamic friction can be mainly attributed to input damping. The characteristic from Fig. 4b is approximated by a quadratic function, namely, Bm( _ 1) = Bm0 sgn( _ 1) + b1 _ 1 + b2 _ 2 (5) 1 where Bm0 constitutes a Coulomb friction torque. The equation of motion for the input shaft for non-zero velocities _ 1 6= 0 can then be readily written as J1 1 + B m + T h = m (6) N while, for vanishing velocities, _ 1 = 0, the motor friction torque Bm becomes a reaction torque and, hence, the equation changes to (7) J1 1 = 0 Bm = m ? Th N for _1 = 0 jBm j jBm0j The equation of motion of the output link is then, J2 2 ? T h = 0 (8) and is not a ected by the change in the model structure being de ned by eqs. (6) and (7), as a result of the di erent character of static and dynamic friction.

3.2 Parameter Identi cation


Equations (1) to (8) de ne the structure of the underlying mechanical model. In the next step, the corresponding parameters must be estimated linking experimental phenomena to modelling assumptions. Ideally, the modelled torque characteristic Th of eq. (4) should b equal the measured torque Th from Fig. 2, for given deformations b b Zi;j = Th( b; _b)i;j ? (Th)i;j = 0 i = 1; :::; n (9)
b where (c denotes measured data, n is the number of data pairs ( b; Th) per measurement, ) and j = 1; :::; m refers to the di erent recorded hysteresis cycles. The unknown parameters are the linear and cubic sti ness coe cients c1; c3 and the hysteresis parameters Bf ]j ; ]j . As shown in Section 2.2, the shape of the measured hysteresis characteristics varies with di erent loading cycles; therefore, the parameters Bf ]j and ]j are considered to depend on the maximum load Th of the respective cycle, for j 2 1; :::; m]. Thus, eq. (9) represents n m nonlinear equations for 2 + 2m unknown parameters pT = (c1 ; c3; Bf ]1; ]1; : : : ; Bf ]m ; ]m ). This overdetermined system of nonlinear equations can be solved by a nonlinear least-square t that minimizes the objective function given below:

3.2.1 Harmonic Drive RH-25

10

Zi;j 2 ! min Z (p) = (10) b p j =1 i=1 max(Th )j b where the scaling factor max(Th)j standardizes the error of the di erent measurements, for j = 1; :::; m. For the parameter estimation of the indirect drive RH-25, n = 10 representative data b points ( b; Th) have been chosen per measurement with j 2 1; :::; 6]. Figure 7 displays the resulting estimated hysteresis functions, that show good qualitative agreement with the b measured data from Fig. 2. The error in the objective function Zi;j = max(Th)j stays within 10%, which furthermore assures a good quantitative transmission representation (Fig. 8).
200

m n " X X

150

100

50

torque [Nm]

50

100

150

200 1

0.8

0.6

0.4

0.2

0 0.2 torsion [deg]

0.4

0.6

0.8

Figure 7: Identi ed hysteresis The estimated parameters for and Bf as functions of the load hystory Th are shown in Fig. 9. The negative slope d =dTh in Fig. 9a preserves the characteristic shape with increasing deformations ( ? ) in eq. (4), while the positive slope dBf =dTh in Fig. 9b represents an increase in friction loss. The dynamic friction torque acting as input damping can also be stated as a nonlinear least-square problem, namely, X _ b _ (11) Z (p) = (Bm( b )i ? Bm( b )i)2 ! min with = (Bm0 b1 b2) whose solution then de nes the estimated torque-velocity function (5) shown as a dashed line in Fig. 4. However, as pointed out before, the non-velocity dependent startup torque
pT
i

11
15

10

relative error [%]

10

15 0

10

20

30 data #

40

50

60

Figure 8: Error in the objective function


400

70 65

350

60
300

55

friction parameter [Nm]


40 60 80 100 120 output torque [Nm] 140 160 180

gamma

250

50 45 40 35 30

200

150

100

25
50 20

20 20

40

60

80 100 120 output torque [Nm]

140

160

180

a) Parameter

b) Parameter Bf

Figure 9: Identi ed parameters and Bf vs. load history Th also depends on the loading cycle. A good approximation of this e ect is obtained by averaging the dissipated work over each operation cycle j , namely,
nP ?1 i=1

Bm0]j =

b j (N bm ? Th)]i ( b1;i+1 ? b1;i)jj

whose values are shown in Fig. 10 as functions of Th .

nP b ?1 j 1;i+1 ? b1;i)jj i=1

(12)

12
0.16

0.14

motor static friction torque [Nm]

0.12

0.1

0.08

0.06

0.04 20

40

60

80 100 120 output torque [Nm]

140

160

180

Figure 10: Startup torque Bm0 vs. load history Th

Kt
Nm 0:2 amp

J1
2:33 10?4 kgm2

N
200

Bm0
0:085 N

b1
5:5 10?4 Nm s rad

b2

c1
2

c3
7:5 106 Nm rad3

?6:3 10?7 Nm s rad2

4870 Nm rad

Table 1: Parameters RH-25 Table 1 summarizes the remaining parameters that are required for the simulation model, i.e., motor torque constant, input inertia and gear ratio, whose values are available from technical speci cations 10], as well as the estimated values for dynamic friction parameters, the sti ness characteristic, and the no-load starting torque.

3.2.2 Harmonic Drive RH-32


The experimental testing also covered a second joint (RH-32) of the REDIESTRO robot. Experimental results for its quasistatic input/output behavior are shown in Figs. 11 and 12 as torsion/output torque and torsion/input torque plots. Since the signature of the measured signal, such as load-history dependent hysteresis and startup torque, is identical to the previously analyzed indirect drive, its modelling and parameter estimation follows the same procedure as outlined in Section 3.1 and Subsection 3.2.1.

13 For the etimation of the hysteresis parameters, n = 10 representative data points have been chosen for each of the m = 4 recorded experiments. The comparison between measured and estimated hysteresis functions from Fig. 11 and 13 shows that the minimization (10), which essentially is a curve tting procedure, mainly lters irregularities in the meab sured data. The relative error in the objective function Zi;j =max(Th)j stays within 15 % (Fig. 14), i.e., it is still reasonably bounded for modelling purposes.
200 200 150 150 100 100 50

output torque [Nm]

50

torque [Nm]

50 50 100 100 150 150 200 0.4 0.5 0.4 0.3 0.2 0.1 0 0.1 torsion [deg] 0.2 0.3 0.4 0.5 0.3 0.2 0.1 0 torsion [deg] 0.1 0.2 0.3 0.4

200

Figure 11: Measured hysteresis

Figure 12: Torque vs. torsion (- - output torque, | input torque)

200

20

150

15

100 10

relative error [%]


0.4 0.2 0 torsion [deg] 0.2 0.4 0.6

50

torque [Nm]

50

5 100 10

150

200 0.6

15 0

10

15

20 data #

25

30

35

40

Figure 13: Identi ed hysteresis

Figure 14: Error in the objective function

The estimated parameters and Bf as functions of the load history Th are shown in Fig. 15. As can be seen, the dependence (Th ) is not necessarily monotonous. The actual friction, instead, arises from the combination of both and Bf , whose resulting frictional e ect gains momentum with increasing load cycles.

14 The velocity-independent, static friction torque is shown in Fig. 16 as a function of Th , while the experimental data for the velocity-dependent, dynamic friction torque can again be well approximated by a quadratic function that is given as a dashed line in Fig. 17.
550 50

500

45

40

friction parameter [Nm]

450

gamma

35

400

30

350

25

300

20

250 20

40

60

80 100 output torque [Nm]

120

140

160

15 20

40

60

80 100 output torque [Nm]

120

140

160

a) Parameter

b) Parameter Bf

Figure 15: Identi ed parameters and Bf vs. load history Th

0.16

o experimental data
0.15

0.25

motor static friction torque [Nm]

motor friction torque [Nm]

0.14

0.2

0.13

0.15

0.12

0.1

0.11

0.1

0.05

0.09 20

40

60

80 100 output torque [Nm]

120

140

160

0 0

50

100

150 200 250 wavegenerator velocity [rad/s]

300

350

400

Figure 16: Startup torque Bm0 vs. load history Th

Figure 17: Dynamic friction (o: experimental data, - -: estimated function)

Table 2 summarizes the technical speci cations as well as the remaining estimated parameters. In this context, it is well worth noting that the measured no-load starting torque Bm0 is 50% smaller than the speci ed catalogue value 10], which emphasizes the variable character of transmission properties even within the same family of harmonic drive gears. This fact suggests that parameters, in general, should be considered to be only valid for one particular harmonic drive setup due to the individual in uence of

15

Kt
Nm 0:2115 amp

J1
7:52 10?4 kgm2

N
260

Bm0
0:0885 N

b1
8:1 10?4 Nm s rad

b2

c1
2

c3
8:8 106 Nm rad3

?7:9 10?7 Nm s rad2

11760 Nm rad

Table 2: Parameters RH-32 manufacturing and assembly imperfections on the overall performance. However, since the underlying structure of the mechanical model does not change, the characteristic parameters of each robotic drive can be readily identi ed by means of an automated parameter estimation scheme, as described in Subsection 3.2.1. Due to the variable magnitudes of the parameters ; Bf , and Bm0, they are also considered to depend on the load history Th . This means, for the model evaluation (4) and (5), that, at each reversal point from loading to unloading, the quantities Th ; ; Bf , and Bm0 are updated according to the current load Th , the three foregoing parameters being linearly interpolated, as indicated in Figs. 9, 10 and 15, 16, respectively. In addition, Coulombtype friction can constrain the actuator rotation, depending on the actuator system state and the balance of acting torques, which causes a change in the model structure given by eqs. (6) and (7). Hence, the system dynamics is governed by di erent equations of motion for actuator velocities _ 1 = 0 and _ 1 6= 0, for torsions _ > 0 and _ < 0, as well as for changing loads Th at the point of reversed relative motion. The key point in the evaluation of such systems with both varying right-hand sides due to hysteresis, as well as changing structure resulting from Coulomb friction at vanishing velocities, lies in the numerical realization of the simulation model. Time-varying constraints and non-smooth transmission torques Th give rise to discontinuities in the system ::: ::: dynamics, i.e., in ( 1; 1; 2), that must be taken into account during the numerical integration. For the complete analysis of the system behavior, we therefore determine the time instants at which an unsteady event occurs, i.e., a constraint or torque transmission function change. These changes in the system description are traced during an integration interval T between two instants, t1 and t2 = t1 + T , by monitoring torque and state variables. Changes in the model structure caused by Coulomb-type friction can be discovered by the following three transition conditions: { A transition from sliding to blocking occurs when the actuator velocity van-

3.3 Model Evaluation

16 ishes within the interval t1; t2], and the static friction torque is bounded, i.e., _ 1(t1) _ 1(t2) < 0 and jBm(t2)j < Bm;0. { Reversed sliding with a discontinuous change in the direction of sliding friction takes place when no static equilibrium exists, i.e., when _ 1(t1) _ 1(t2) < 0 and jBm(t2)j = Bm;0 holds. { A static friction constraint is broken o when the transition from _ 1(t1) = 0, jBm(t1)j < Bm;0 to jBm(t2)j > Bm;0 is ful lled. In contrast to this structure-variant friction constraint, that is controlled by the actual saturation of the static friction force Bm (t2), changes in the hysteresis representation depend solely on the system state history, i.e., the torque transmission function changes at each reversal point from loading to unloading and, hence, whenever the condition _ (t1) _ (t2) < 0 is ful lled. The time histories of the magnitudes _ 1; Bm, and _ then indicate each change in the system description whose occurrence at time t 2 t1; t2] is iteratively determined by interval bisection T=2 during the integration. Thus, the simulation task is split into a series of initial value problems where the underlying di erential equations have a constant structure, and the step size control is such that discontinuities always occur at the boundary t of the integration interval t1; t ] and t ; t2], respectively. The design of the software structure on integrator level including variable step size control is described in detail in 11, 12]. It is worth noting that the model presented for a joint transmission unit can be easily incorporated in the robot dynamics equations. For an f -degree-of-freedom manipulator, the equations of motion are then given by M ( ) + h( ; _ ) = R (u ? B m ) + T (13) T ; T )T 2 IR2f =( M L where M is the matrix of generalized inertia, h is a vector function that contains gravitational, Coriolis, and centrifugal forces, Ru is the control input, and M , L represent the motor and link coordinates. The vector function T , with components Ti = ?Th;i=Ni ; Ti+f = Th;i, takes into account the e ect of compliance and hysteresis between motor shaft and outer link, while RBm is the friction torque acting on the motor shaft. In case of static friction in joint i, the friction torque Bm;i becomes a reaction torque for vanishing actuator motion M;i = 0. The dynamics of a constrained actuator, however, is always decoupled from the remaining system dynamics, i.e., the actuators cannot form kinematic loops with each other, so that each joint module can be treated separately, as described in Section 3.1.

17

4 Model Veri cation and Results


The previous model development is based on the inverse problem of system analysis, i.e. the system behavior is described according to given input/output relationships. More challenging questions arise when the model is used to forecast output signals for given input signals. In order to verify the quality of the model structure and its parameters, a diagnostic veri cation has to be performed that involves the dynamic excitation of the system and the comparison of actual output signals to numerically derived data. For this testing, the output shaft was mounted to ground, i.e., 2 = 0, and the motor has been driven by a sinusoidal current input with varying amplitudes and frequencies.

A comparison of experimental and simulated transfer functions, de ned by maximum angular input rotation per input current, is shown in Fig. 18 for the harmonic drive RH-25. The results show a slight reduction in the resonance frequency of the system with increasing excitation amplitudes indicating a small variation of the e ective sti ness depending on operation conditions. Furthermore, the gain 1=im increases with higher excitation amplitudes, clearly indicating the nonlinear behavior of the system under consideration.
1.8 1.6 1.4 1 amp 1.5 amp . 2 amp

4.1 Results for the Harmonic Drive RH-25

transfer function [rad/amp]

1.2 1 0.8 0.6 0.4 0.2 0 0

6 f [Hz]

10

12

Figure 18: Transfer function (o for im = 1; 1:5; 2 amp)

o: experimental results, |, - -, -.- : simulation results

For excitation frequencies f < 2:5Hz, the system response contained stick-slip cycles that disappeared at higher frequencies because of the increasing phase shift between m

18 and ( 1; Th), so that the torque m ? Th=N applied to the input shaft always exceeded the Coulomb friction torque during changes in the wave-generator direction of motion. With increasing vibration amplitudes 1, the torque Th becomes dominant and the resulting motion approaches a sinusoidal signal (Fig. 19).
1 0.8 0.6 0.4

input angle [rad]

0.2 0 0.2 0.4 0.6 0.8 1 0

0.5

1.5 time [s]

2.5

Figure 19: Simulated time response (im = 1 amp, | f = 1Hz,

f = 4Hz)

100 80 60

input angular velocity [rad/s]

40 20 0 20 40 60 80 100 4

0 1 input angle [rad]

Figure 20: Phase plot for f = 5Hz, im = 1; 1:5; 2 amp (| simulation,

measurement)

19 Figure 20 displays a comparison between measured and simulated phase-plane plots in the resonance region f = 5Hz that show good agreement for di erent excitation amplitudes. Further simulation runs, where individual friction and damping values were set to zero, revealed that the hysteresis in the harmonic drive produces the most damping; startup torque also strongly in uences the gain, while velocity-dependent friction contributed the least damping.

4.2 Results for the Harmonic Drive RH-32


The dynamic testing with harmonic excitation has also been conducted for the second robotic joint. Figure 21 contrasts experimental transfer functions for two di erent current amplitudes with results obtained by simulation. As expected, the gain 1=im increases with larger excitation amplitudes. This e ect results from the Coulomb friction torque that reduces the net input torque m ? Bm;0 in the undercritical excitation region. For increasing motor torques m;2 > m;1, the ratio of net torques ( m;2 ? Bm;0)=( m;1 ? Bm;0) is greater than the ratio of input torques m;2= m;1 , which contributes to the nonlinear behavior shown in Fig. 21.
1.5 1 amp 2 amp

transfer function [rad/amp]

0.5

0 0

4 f [Hz]

Figure 21: Transfer function (o o: experimental results, |,- - : simulation results for im = 1; 2 amp) For an estimation of the overall damping, consisting of dynamic friction, static friction and hysteresis loss, we can approximate the system dynamics by a linear second order

20 system that, in normalized form, appears as Kt imsin(!t) 2 (14) 1 + 2 ! n _ 1 + !n = J1 with !n = p !d 2 1? and represents an equivalent damping ratio, while !n and !d are the natural frequency and the damped frequency, respectively. The experimental transfer function from Fig. 21 can be best approximated in the resonance region with values = 0:32 and !n = 24:5 1=s2, for im = 1 amp, while = 0:35; !n = 18:5 1=s2, for im = 2 amp. The comparison with the damping ratio = b1=(2J1!n ) that results from linear viscous damping (Table 2) and lies in the neighbourhood of 0:02 0:03 shows that damping is mostly produced through motor friction and hysteresis loss in the tooth engagement area. Furthermore, the absolute values of the equivalent damping ratios illustrate that the vibrations of the actuator are heavily damped. This fact can also be observed in experiments where the actuator is excited by a step input current that does not produce signi cant overshoot or vibrations in the transient system response (Fig. 22).

0.6

0.5

0.4

input angle [rad]

0.3

0.2

0.1

0 0

0.1

0.2

0.3 time [s]

0.4

0.5

0.6

Figure 22: Step input response, im = 1:1 amp ( : experimental results, |: simulation results)

21

5 Conclusions
In this paper, we analyzed the nonlinear torque transmission of indirect drives with harmonic drive gears, commonly used in robotic systems. First, we study the indirect drive input/output signals that are derived from independent quasistatic and stationary dynamic experiments. Based on the signals signature, we then develop a mechanical model that takes into account compliance and friction (hysteresis) in the harmonic drive transmission, the Coulomb-type friction at the input shaft, that is required to get the motor into motion, as well as viscous-type damping that depends nonlinearly on the input angular velocity. A mathematical model is formulated, its parameters being estimated by a nonlinear least-square t. One important requirement in this context is that the experimental data cover the whole operation range of the harmonic drive transmission so that the parameters that depend nonlinearly on the load history can be approximated su ciently well by linear interpolation. The quality of the simulation model is then veri ed by exciting both the testbed and the model with a sinusoidal input signal of varying amplitude and frequency. Two robotic joints, that show the same structural input/output behavior, have been analyzed and the comparison of the output signals shows a good agreement between model and experiment. The dynamic system repsonse features heavy damping, which is produced mostly by hysteresis losses in the harmonic drive gear and Coulomb-type motor friction, while the in uence of velocity-dependent damping is less signi cant. The presented transmission model has been devised in such a way that in can be easily incorporated into the framework of multibody system dynamics in order to facilitate the modelling and control of the complete robot. Thus, the model can be used for the design of robust motion/force control schemes; it can also be used for the prediction of controller performance via simulation, which takes into account constraints on the actuator motion, due to Coulomb friction.

22

References
1] Ranjbaran, F., Angeles, J., Gonzalez-Palacios, M.A. and Patel, R.V.: The mechanical design of a seven-axes manipulator with kinematic isotropy, to appear in the Journal of Robotics and Intelligent Systems, 1994. 2] Tuttle, T. and Seering, W.: Kinematic Error, Compliance, and Friction in a Harmonic Drive Gear Transmission. ASME Advances in Design Automation, Vol. 65-1, 1993, pp. 319 { 324. 3] HDC Cup Component Gear Set Selection Guide. Harmonic Drive Technologies, Peabody, MA. 4] Legnany, G. and Faglia, R.: Harmonic Drive Transmissions: the e ects of their elasticity, clearance and iregularity on the dynamic behaviour of an actual SCARA robot. Robotica, Vol. 10{4, 1991, pp. 369{376. 5] Marilier, T. and Richard, J.A.: Non-Linear Mechanic and Electric Behaviour of a Robotic Axis With a Harmonic-Drive Gear. Robotics and Computer-Integrated Manufacturing, Vol. 5, No. 2-3, 1989, pp. 129 { 136. 6] Kircanski, N., Goldenberg, A.A. and Jia, S.: An Experimental Study of Nonlinear Sti ness, Hysteresis, and Friction E ects in Robot Joints with Harmonic Drives and Torque Sensors. Preprints of the Third International Symposium on Experimental Robotics, Oct. 28 { 30, 1993, Kyoto, pp. 147 { 154. 7] Tuttle, T.: Understanding and Modeling the Behavior of a Harmonic Drive Gear Transmission. MIT Master Thesis, Technical Report No. 1365, 1992. 8] Maghzal, A.J.: Characterization of Harmonic Drive Parameters. Technical Report 305-463N. Centre for Intelligent Machines, McGill University, 1994. 9] Laschet, A.: Simulation von Antriebssystemen. Springer-Verlag, 1988. 10] Harmonic Drive Gearing. Harmonic Drive Systems, Inc., Hauppauge, NY. 11] Sey erth, W. and Pfei er, F.: Modelling of Time-Varying Contact Problems in Multibody Systems. Proc. of 12th Symposium on Engineering Applications of Mechanics, June 27 { 29, 1994, Montreal, pp. 579 { 588. 12] Sey erth, W.: Modellierung unstetiger Montageprozesse mit Robotern. Fortschrittberichte VDI, 11, No. 199, VDI-Verlag, Dusseldorf, 1993.

Das könnte Ihnen auch gefallen