IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO.

3, JUNE 2012

555

Real-Time Identication of HuntCrossley Dynamic Models of Contact Environments

Amir Haddadi, Student Member, IEEE, and Keyvan Hashtrudi-Zaad, Senior Member, IEEE
AbstractReal-time estimates of environment dynamics play an important role in the design of controllers for stable interaction between robotic manipulators and unknown environments. The HuntCrossley (HC) dynamic contact model has been shown to be more consistent with the physics of contact, compared with the classical linear models, such as KelvinVoigt (KV). This paper experimentally evaluates the authors previously proposed singlestage identication method for real-time parameter estimation of HC nonlinear dynamic models. Experiments are performed on various dynamically distinct objects, including an elastic rubber ball, a piece of sponge, a polyvinyl chloride (PVC) phantom, and a PVC phantom with a hard inclusion. A set of mild conditions for guaranteed unbiased estimation of the proposed method is discussed and experimentally evaluated. Furthermore, this paper rigorously evaluates the performance of the proposed single-stage method and compares it with those of a double-stage method for the HC model and a recursive least squares method for the KV model and its variations in terms of convergence rate, the sensitivity to parameter initialization, and the sensitivity to the changes in environment dynamic properties. Index TermsDynamic model identication, HuntCrossley (HC), KelvinVoigt (KV), online parameter estimation.

I. INTRODUCTION

OBOTIC tasks often involve continuous or intermittent contacts between robots and various environments. The interaction between a slave robot and body tissues in robotassisted minimally invasive surgeries [1], [2], a robotic nger grasping an object [3], a teleoperated excavator bucket during remote excavation [4], and the interaction between foot and ground during the locomotion cycle of a walking machine [5] are a few examples of contact tasks. Real-time estimates of contact dynamics have been used for the design of indirect and model reference adaptive controllers for stable contact in robotic and telerobotic applications [6][9].

Manuscript received October 18, 2010; revised April 29, 2011 and October 25, 2011; accepted December 22, 2011. Date of publication February 7, 2012; date of current version June 1, 2012. This paper was recommended for publication by Associate Editor T. Murphey and Editor B. J. Nelson upon evaluation of the reviewers comments. This work was supported in part by Natural Sciences and Engineering Research Council of Canada. A. Haddadi was with the Department of Electrical and Computer Engineering, Queens University, Kingston, ON K7L 3N6, Canada. He is now with the University of British Columbia, Vancouver, BC V6T 1Z4, Canada (e-mail: haddadi@ieee.org). K. Hashtrudi-Zaad is with the Department of Electrical and Computer Engineering, Queens University, Kingston, ON K7L 3N6, Canada (e-mail: khz@queensu.ca). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TRO.2012.2183054

Available methods that are widely used to model the environment in real-time control applications are mainly limited to linear KelvinVoigt (KV) models, in which the relationship between the penetration of contacting bodies and the resulting force is modeled by a parallel connection of a linear spring and a linear damper. Linear models have been identied in real time using estimation methods such as recursive least squares (RLS) and its variations such as exponentially weighted recursive least squares (EWRLS) [7], [10], [11]. However, the KV or linear massdamperspring models, in general, are shown to have physical inconsistencies in terms of power exchange during contact. This shortcoming, which results in negative contact force during rebound, is more visible in dynamic models of soft environments, such as human body tissues [12]. Therefore, the estimated forces and dynamic parameters using linear contact models may not be suitable for control of contact tasks. In 1975, Hunt and Crossley [13] showed that a positiondependent environment damping model is more consistent with physical intuition. It is further shown that the HuntCrossley (HC) model is consistent with the notion of coefcient of restitution, representing energy loss during impact [14]. Therefore, such a nonlinear model can potentially improve the estimation of force and dynamic parameters, which, by itself, will improve the performance of many robotic, haptic, and telerobotic tasks. However, fast and accurate identication of the HC nonlinear models remains a challenge and severely limits the use of this model for real-time applications. Diolaiti et al. proposed a bootstrapped double-stage method for online identication of the HC model [12]. However, this method is sensitive to parameter initial conditions as demonstrated by simulations in [15] and by experimental results in this paper. Moreover, due to the propagation of error from one stage to the next, the double-stage method suffers from slow parameter convergence. Recently, Haddadi and Hashtrudi-Zaad have proposed a novel single-stage method for online estimation of the HC model [15]. They proved estimation consistency (unbiased estimation) and provided mild conditions upon which the method is applicable. The single-stage method has been used in [16] to improve model-mediated teleoperation systems and in [17] for stiffness modulation in haptic augmented reality applications. Although the single-stage method has been simulated in [15], it has not been rigorously evaluated with experiments. In this paper, we experimentally evaluate the previously proposed single-stage estimation method for different types of environment contact materials. We investigate the sensitivity of the method to initial conditions and model parameter variations and study the estimation convergence rate and force prediction accuracy in comparison with the double-stage method applied to an HC model

1552-3098/$31.00 2012 IEEE

556

IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO. 3, JUNE 2012

and an RLS variant applied to a KV model. Then, we provide intuitive insights on how to design a robotic task for an unknown environment in order to benet from efcient identication. The remainder of this paper is organized as follows. The linear KV and the nonlinear HC contact models are presented in Section II. The double-stage and single-stage online identication methods are described in Section III. Experimental results comparing the performance of the two methods are presented and analyzed under a variety of conditions for three different contact environments in Section V. The effect of various input excitation and parameter variation during intermittent contact in an automated tissue property estimation task are studied in Sections VI and VII, respectively. Section VIII draws conclusions and provides suggestions for future work. II. CONTACT DYNAMICS MODELS A. KelvinVoigt Linear Contact Model The most common environment dynamic model for robotic applications is the linear KV model, which incorporates the dynamics of a linear damperspring system F (t) = Kx(t) + B x (t), 0, x(t) 0 x(t) < 0 (1)

Fig. 1.

Double-stage identication method for the HC model [12].

to linearize the nonlinear model so that all three parameters of the HC model can be estimated at the same time in a single stage. This makes the real-time identication process faster and computationally more efcient [15]. In this section, we provide a brief overview of the doublestage parameter estimation method proposed by Diolaiti et al. [12] and the single-stage identication method proposed by the authors in [15] and discuss their differences. A. Double-Stage Identication Method Fig. 1 illustrates the double-stage parameter estimation method [12]. In this method, the estimation of the dynamic parameters (Kc , Bc ) is partially decoupled from the parame , the following linear ter n. In stage 1 , assuming a known n dynamic equation is used to estimate Kc and Bc :
(t) (t) F (t) = Kc [xn ] + Bc [xn x (t)] + 1

where x(t) represents the robot penetration inside the environment, F (t) represents the contact force, and K and B denote the environment stiffness and damping, respectively. The dynamic parameters of this model can be easily estimated using linear system identication techniques, such as variations of least squares. The KV model displays both physical limitations and intuitive inconsistencies [15], [18], which will be discussed later through experiments. B. HuntCrossley Nonlinear Contact Model Nonlinear models have been shown to be in better agreement with the dynamic behavior of physical environments [18]. Specically, human biological tissues have been reported to show nonlinear behavior [2]. In order to overcome the problems of the KV linear model, Hunt and Crossley [13] proposed the following nonlinear model: F (t) = (t), Kc xn (t) + Bc xn (t)x 0, x(t) 0 x(t) < 0 (2)

(3)

where 1 is the error generated from using n instead of n. In c and B c parameters, the parameter stage 2 , assuming known K n is estimated according to log F (t) = n[log x(t)] + 2 c + B c x K (t) (4)

in which the viscous force depends on contact penetration. Here, denotes Kc xn denotes the nonlinear elastic force, and Bc xn x the nonlinear viscous force. The parameter n is typically between 1 and 2, depending on the material and the geometric properties of contact. III. LINEAR IDENTIFICATION OF THE NONLINEAR HUNTCROSSLEY MODEL The nonlinear nature of HC models is intuitively consistent with the physics of contact; however, the resulting computational complexity of the double-stage identication method has restricted the use of the HC model in real-time robotic applications. Therefore, the authors proposed a different approach

c and where 2 is the error resulting from the estimation of K Bc . For the unbiased estimation of all three parameters, both 1 and 2 must be zero-mean stochastic processes. Although the convergence of this method has been analyzed [12], the proof of estimation consistency has been provided under three conditions that may substantially limit the applicability of this method. 1) In order for 1 to be a zero-mean stochastic process, it is assumed that n = n n is always small such that 1 x n log x n = n log x. In other words, the approximation holds only when the condition x n 1 is met. This condition cannot be realized at the beginning of the estimation process, where x is very small and n is potentially large. For instance, for the initial error n = 0.2 and 1-cm penetration, (0.01)0 . 2 = 0.398 which is not close to 1. 2) In order to have an unbiased estimation in 2 , the following necessary condition must be satised: Kc + Bc x 1 xn (5)

HADDADI AND HASHTRUDI-ZAAD: REAL-TIME IDENTIFICATION OF HUNTCROSSLEY DYNAMIC MODELS OF CONTACT ENVIRONMENTS

557

c Kc and B := B c Bc . This conwhere Kc := K dition is met at the initial estimation period where x is small, as well as after convergence when Kc and Bc become small. However, if the parameters of the system change during operation, condition (5) is no longer met, and there is no guarantee for 2 to be unbiased. 3) The statistical condition )xn ] = 0 E [n(Kc + Bc x (6)

must be satised, which is the case if the two estimators converge independent of each other. However, the convergence of 2 is dependent on 1 and vice versa, as each estimator relies on the resulting estimates from the other estimator. B. Single-Stage Identication Method As discussed in Section III-A, the applicability of the doublestage parameter estimation method is limited under certain conditions. These limitations may result in inconsistent estimations due to the choice of initial conditions that are far from the actual unknown values. Therefore, the authors provided a different method to linearize the nonlinear HC model so that all three model parameters can be identied in one stage during a real-time process. It has been shown in [15] that by taking the natural logarithm of both sides of the HC model (2) for x 0, we obtain ln[F (t)] = ln[Kc xn (t)(1 + (t) Bc x )] + Kc Kc xn (t) (t) Bc x ] + Kc Kc xn (t) (7) where includes the modeling error and the measured noise during the process. As we know for the natural logarithm, ln(1 + ) 1. Therefore, assuming = for || Bc x Bc x (t) (t) + + Kc Kc xn (t) Kc Kc xn (t) (7) can be rewritten as ln[F (t)] = ln(Kc ) + Bc . (9) x (t) + n ln[x(t)] + Kc Kc xn (t) 1 (8)

is a rather conservative value in many cases. Therefore, being marginally close to this value, or even slightly violating this condition, does not necessarily mean that the identication is not valid. Since the values of Kc and Bc are unknown in realtime experiments, their estimates can be used as an alternative to validate condition (10), as will be seen in Section V. With regard to the feasibility of condition (10), in many practical applications, the speed of operation within contact is not high and condition (10) is often met. However, it should be noted that the speed of operation is not fully controlled by the user as it depends on the characteristics of the desired task and the capabilities of the robot. Therefore, the fact that condition (10) is not guaranteed can be considered a shortcoming of this method. With regard to the second term of approximation (9), depending on the power of noise and the type of environment, i.e., Kc , a reasonable minimum penetration must be chosen so that | K c xn ( t ) | is small enough to satisfy (8). The identication process must be stopped when the penetration is smaller than this threshold. Considering the aforementioned conditions, the linearized system (9) can be identied using the least squares family of estimation methods. To this purpose, the environment dynamics (9) is linearly parameterized as k , yk = T k k + T k xk > 0 (11)

where is the regressor vector, k is the vector of dynamic k parameters, and k = m k + K n represents modeling error c xk and force measurement noise at the sample time t = k.T , where k is the iteration number, and T is the sampling time. In addition k , ln(xk )], = ln(Kc ), T k = [1, x Bc ,n Kc
T

= ln(Kc ) + n ln[x(t)] + ln[1 +

, yk = ln(Fk ).

(12) Among variations of the RLS methods, EWRLS is an estimation method that is more suitable for environments with variable dynamic properties. The EWRLS update equations can be written as [19] Lk +1 = Pk +1 = k +1 Pk k +1 + T k +1 Pk k +1

1 [Pk Lk +1 T k +1 Pk ] k + Lk +1 [Fk +1 T = k +1 k ]

(13)

x (t) The assumption in (8) implies that | B cK | and | K c xn ( t ) | are c

negligible. To satisfy sider the condition

x (t) | | B cK c

1, as a rule of thumb, we con< 0.1Kc Bc (10)

where P is the covariance matrix, and is the forgetting factor. When = 1, the RLS method is achieved. At every sampling time, the estimated parameters of the model are derived according to
k (1) k (1) k (3). ck = e ck = e K , B k = k (2), n

with the approximation threshold of 0.1. This threshold would result in a minimum uncertainty of 0.005 in (8), which is small and acceptable, considering the potential magnitude of ln(Kc ) and n ln[x(t)] in (9). Although the effect of the threshold on the identication process depends on the other terms in (9), such as the value of Kc , the force, and the amount of measurement noise, our experiments and simulations have shown that 0.1

The algorithm should check for singularities in the logarithmic functions that may occur during the operation. As in every parameter estimation method, the estimation convergence relies upon the persistency of excitation (P.E.) of the robot end-effector trajectory. Since there are three parameters to be identied, as a rule of thumb, a combination of two sine waves would be sufcient [20]. In practical applications, this excitation condition is

558

IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO. 3, JUNE 2012

Fig. 3.

Desired trajectory perpendicular to the contact surface.

Fig. 2. Experimental setup consisting of a twin pantograph robotic manipulator equipped with a Nano-25 force/torque sensor and a poking device.

Fig. 4. Three different contact materials identied through experiments. From left to right: rubber ball, sponge, and PVC phantom.

not guaranteed to be met; however, the nonlinearity of the function ln() can make the identication signal richer [20]. The P.E. condition will be experimentally investigated in Section VI. The proof of unbiased estimation for the single-stage identication method can be found in [15]. In the next sections, we will experimentally evaluate the single-stage method and demonstrate the advantages of this method over the double-stage method for the HC model and the EWRLS method applied to the KV model. IV. EXPERIMENTAL SETUP FOR REAL-TIME IDENTIFICATION Fig. 2 shows a picture of the experimental setup consisting of a Quanser planar twin pantograph robotic manipulator equipped with an ATI Nano-25 force/torque sensor and a poking device. The twin pantograph is a redundant robot that consists of two pantographs, each directly driven by two dc motors at the base joints. The angle of rotation of each motor is measured by a high-resolution encoder with 20 000 counts/rev. The position of the end-effector is computed from forward kinematic relations. The contact force is measured at a resolution of 1/8 N. The robot is position controlled to follow a desired trajectory within the contact material. The trajectory, which will be discussed later, consists of various sinusoidal components. The control system and the online identication algorithms are implemented using MATLAB RTW Toolbox and Quanser QuaRC 1.1 real-time system operating at a sampling rate of 1 kHz on a 2.4-GHz Quad CPU. A. Desired Trajectory for the Probe Different trajectory commands for the robot are used in order to determine the least possible level of excitation for identication. First, the following combination of three sinusoidal signals is chosen as the desired trajectory in the direction perpendicular to the contact surface x(t) = 1 sin(4t) + 1.8 sin(11t) + 1.8 sin(15t) + x0 (mm).

The frequency and coefcients are chosen to ensure suitable penetration inside the environment as well as the richness of the excitation input. The bias x0 is added to push the robot smoothly inside the environment about 10 mm such that the robot remains inside the object during the entire contact task. A sample position trajectory for the rst 10 s is shown in Fig. 3. A 2-s delay in the position command is implemented to measure and remove any bias in the force sensor measurements before the actual experiment starts. B. Contact Material Three different contact materials (environments) are used for the experiments. They include a rubber ball, a piece of sponge, and a polyvinyl chloride (PVC) phantom. Each environment displays a specic behavior that results in important conclusions about the identication method and the advantage of the nonlinear HC model over the linear KV model. Fig. 4 illustrates the experimental setup with the robot in contact with the three different environments. V. EXPERIMENTAL PROCEDURE AND RESULTS A. Environment: Rubber Ball An elastic rubber ball is used as the contact material for the rst set of experiments as shown on the left side of Fig. 4. This type of contact, when compared with the other two, displays a dominant elastic behavior, which returns the probe quickly to its original state once the stress is removed. 1) Initial Conditions: To investigate the sensitivity of the identication algorithms to parameter initialization, three experiments are conducted for each environment using small and large initial conditions for the identied parameters. The choice of small values for Kc 0 , Bc 0 , and n0 is trivial. The set of high values is established as two or three times the average of the parameters in the steady state obtained by performing the identication experiments with the small initial values. To estimate the parameters of the HC model, we examine the small values for Bc 0 and Kc 0 : once with n0 = 1 and once with the maximum

HADDADI AND HASHTRUDI-ZAAD: REAL-TIME IDENTIFICATION OF HUNTCROSSLEY DYNAMIC MODELS OF CONTACT ENVIRONMENTS

559

TABLE I RUBBER BALL: PARAMETER INITIAL VALUES FOR HC AND KV MODELS

Fig. 7. Effect of on the identied parameters of the HC model using the single-stage method for one experiment with similar initial conditions.

Fig. 5. Rubber Ball. Estimated parameters of the HC model using the singlestage method for three repetitions of the real-time experiment using the rst set of initial conditions.

Fig. 6. Convergence condition of the single-stage HC environment contact identication method for the rubber ball. For reliable estimated parameters, we 1K c should have x < 0 .B . c

expected value n0 = 2. The two experiments help us to focus on the effect of n. In the third experiment, we change Bc 0 and Kc 0 to their larger values to focus on the effect of Bc and Kc . The three sets of initial values for Kc 0 , Bc 0 , and n0 for the rubber ball are shown in Table I.1 Since the KV model has only two parameters, i.e., K and B , two experiments with low and high values, as summarized in Table I, are considered. 2) Validity of the Single-Stage Method: The results of the estimation of the HC model parameters using the single-stage method are shown in Fig. 5. The gure illustrates the identied parameters for three repetitions of the real-time experiment with the rst set (set I) of initial conditions for all trials. The agreement of the estimated parameters veries the experimental results and points at the consistency of the single-stage method. As an alternate method to check the applicability of the singlestage method and the validity of its results, the approximation condition (10) is investigated. Since the correct values of Kc and Bc are not available, their estimates are used. Fig. 6 comc pares |x | and 0.1 K c for a set of collected data and a set of initial B conditions over the entire period of time during which the probe is in contact with the ball. The gure shows that the condition 1K c x < 0 .B is satised and the identication results are valid c after approximately 90 ms from the start of the experiment. 3) Effect of Forgetting Factor and Covariance Matrix on the Single-Stage Method: The identication process for the
1 For singularity issues, K c 0 = 0 cannot be used in our identication process, and instead, K c 0 = 0 .1 is used.

previous three trials starts at t = 2 s with the initial forgetting factor 0.99, which nearly reaches unity in one second according to the exponential relationship = 1 0.01exp(5(t 2)). The choice of exponential forgetting factor is motivated by the fact that for constant = 1, as in RLS, the estimation converges more slowly, while for a constant below unity, instability occurs. The forgetting factor is chosen based on the type of environment under examination. For contacts with varying parameters, a lower forgetting factor is selected, whereas for contacts with constant parameters, a value closer to unity is selected. The choice of the time constant in the exponent also has signicant effect on the convergence rate and convergence stability. Fig. 7 shows the slow convergence experienced when was chosen as unity. Our experiments have shown that using a large time constant to bring close to unity results in instability. Using = 1 0.01exp( (t 2)), with 5 8, creates a balance between the convergence rate, sensitivity to noise, and estimation stability. Therefore, in this paper, we use = 1 0.01exp(5(t 2)) for the forgetting factor in order to balance estimation speed and estimation convergence. Fig. 7 illustrates the parameters that are estimated in real time after performing the single-stage identication method on one set of experimental data with different proles. It is clear that by using the RLS method, i.e., = 1, the estimation becomes very slow. The time constant of 10 s, i.e., = 0.1, results in severe uctuations and instability, whereas = 5 provides the best performance. While the initialization of the covariance matrix P is expected to affect the speed of convergence at the beginning of identication, the experiments show that matrices with norms larger than 10 result in similar convergence rates. On the other hand, it has also been approved that the larger the norm of P , the larger the magnitude of the initial overshoots of the estimated parameters. It has further been observed that the extent of these jumps mainly depends on the initial values of the estimated parameters. The effect of the initial conditions will be discussed next. 4) Effect of Initial Conditions on the Two HuntCrossley Identication Methods: Fig. 8 shows the online estimated parameters and their corresponding force prediction errors for the various initial conditions listed in Table I. The results in Fig. 8 are related to the three discussed online identication methods (HC: single stage, HC: double stage, and KV), which are all applied to the same set of collected data. The percent relative root-mean-square error (RMSE%) is also provided in

560

IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO. 3, JUNE 2012

Fig. 8. Rubber Ball. Estimated parameters and the force prediction errors for various initial conditions for HC and KV models. The KV model has one fewer parameter to display. TABLE II RUBBER BALL: RMSE% OF THE HC AND KV MODELS

Table II. Since the KV model has no parameter n to identify, its corresponding RMSE% value for the initial set II is not available. It is clear from the results that the single-stage method outperforms the double-stage method in terms of convergence rate,3 i.e., speed of convergence at the beginning of the identication process, force prediction error, and consistency of estimation. Moreover, the single-stage method is capable of identifying the environment dynamic parameters for a wide range of initial conditions, whereas the double-stage method lacks this capability. The reason for such inconsistency using this set of initial conditions is that the convergence conditions for the double-stage method are not satised, as discussed in Section III-A. Fig. 9 shows the identication results obtained from the double-stage method for four different initial conditions that are selected within the vicinity of the nal values. The double-stage method produces consistent and converging results only for a small range of initial conditions, which is not desirable for applica2 In order to exclude the effect of large force prediction error experienced in the rst few milliseconds of contact, the calculation of %RMSE for all the methods are performed after 0.2 s of contact. 3 Convergence time is dened as the time required for the estimates of a parameter for all initial conditions to converge within a 5% difference.

Fig. 9. Rubber Ball. Estimated parameters of the HC model using the doublestage method for different initial conditions.

tions in which limited or no information is available about the contact material. 5) Comparing the Identied HuntCrossley and Kelvin Voigt Models: Here, we again focus on the results in Fig. 8 and Table II to compare the HC model obtained from the singlestage method and the identied KV model. The larger prediction error for the KV model implies that the HC model better represents the physical properties of the object. Although parameters converge for both models, the convergence time is shorter for the single-stage method. Considering that the same measurements, i.e., position and force, and the same initial conditions for the EWRLS identication processes have been utilized in both models, the difference in convergence rate may, then, be related to the role of the natural logarithms of position and force in the single-stage method in making the identication excitation richer and resulting in faster estimation. As previously mentioned, one potential shortcoming of KV models is the negative predicted force when the probe reaches close to the environment surface as it moves in the outward direction at a high velocity. In this case, i.e., x(t) 0, and since x (t) is negative and large, a negative force is predicted by the

HADDADI AND HASHTRUDI-ZAAD: REAL-TIME IDENTIFICATION OF HUNTCROSSLEY DYNAMIC MODELS OF CONTACT ENVIRONMENTS

561

Fig. 10. (Left) Position prole and (right) the measured and predicted force position hysteresis loop for the KV model. TABLE III SPONGE: PARAMETER INITIAL CONDITIONS FOR THE HC AND KV MODELS

Fig. 11. Sponge. Estimated parameters of the HC model using the single-stage method for three repetitions of real-time experiment with the rst set of initial conditions.

model. However, in reality, the force exerted on any environment is always positive during the compression and rebound phases. Such behavior leads to power transfer from the robot to the environment during the restitution phase, which is in contrast with human intuition. This problem primarily occurs in systems with higher damping at high speeds. In order to demonstrate this behavior, a simple experiment is performed on the rubber ball with a penetration prole, which comes close to zero at certain periods in time. Fig. 10 illustrates the superimposed forceposition hysteresis loops obtained from actual measurements and the estimated model. To avoid the negative force effect in all experiments described throughout the paper, with the exception of the aforementioned results, we have considered a sufciently large x0 . This allows us to focus on comparing the HC and KV models to predict contact behavior once the probe is inside the environment. B. Environment: Sponge In this section, a thick piece of sponge, as shown in Fig. 4, is used as the contact material for dynamic identication. The dynamic response of a sponge to contractions is different compared with that of the ball. Because of the larger damping property of the sponge, it requires substantially larger time for restitution. In this section, we experimentally identify and compare the HC models, obtained from the single- and double-stage methods, with the KV model for this dynamic property of the sponge. 1) Initial Conditions: Three sets of initial values for Kc , Bc , n, and two sets of initial values for K and B , as shown in Table III, are considered for the estimation of the dynamic parameters of the sponge. The selection of the initial values follows the strategy explained in Section V-A for the rubber ball. The results of real-time estimations are presented next. 2) Validity of the Single-Stage Method: Fig. 11 shows the estimated HC parameters of the sponge using the single-stage identication method for three trials with the Set I initial conditions described in Table III. One property that differentiates the sponge from the elastic rubber ball is the damping-related

Fig. 12. Sponge. Convergence condition of the single stage HC environment contact identication method. For reliable estimated parameters, the condition 1K c should be met. x < 0 .B c

parameter Bc of the sponge, which is about ten times higher than its corresponding Kc value. In order to conrm the applicability of the single-stage method for the sponge environment, the approximation conc dition (10) is examined. To this purpose, |x | and 0.1 K c are B compared in Fig. 12 for one set of collected data and one set of initial conditions over the entire period of time that the probe is in contact with the sponge. The gure shows that condition (10) is violated for the rst 5 s of contact. However, from that point on, condition (10) is met with a small margin. The sponge is less elastic and more damped than the rubber ball, resulting in substantially lower value for Kc /Bc or a lower boundary on the velocity. As a result, for faster operations, i.e., higher velocity, the identication results lose their accuracy and consistency. Therefore, as previously discussed, one of the main shortcomings of the single-stage method is that the convergence conditioners may not be met for a highly damped environment, especially for the tasks in which the speed of interaction is not fully controlled by the user. For other tasks for which the speed of operation can be controlled, such as palpation for physical examinations, a lower speed of operation is recommended. 3) Effect of Initial Conditions and Comparison Between Different Methods: In order to determine the effect of initial conditions on the identication of the HC model using the single-stage and the double-stage methods, as well as the linear identication of the KV model, different initial conditions, as listed in Table III, are applied to one set of collected data. Fig. 13 shows the estimated parameters and the prediction error proles for the three methods, and Table IV summarizes the RMSE% values. The results show that the speed of parameter convergence for the EWRLS applied to the single-stage method for HC model is the same as that of the KV model and is not sensitive to the large changes in the initial conditions. In contrast, the doublestage method shows inconsistent results for such a large range of

562

IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO. 3, JUNE 2012

Fig. 13. Sponge. Estimated parameters and the force prediction errors for various initial conditions for the HC and KV models. The KV model has one fewer parameter to identify and display. TABLE IV SPONGE: RMSE% OF THE HC AND KV MODELS

initial conditions. This highlights the advantage of the singlestage method over the double-stage method in identifying the contact dynamic parameters for a wide range of initial conditions. Our experiments show that the double-stage method can only converge when the initial conditions are within 10% of their nal values. In terms of model force prediction, a larger force prediction error is observed for the identied KV model of the sponge than the HC model that uses the single-stage method. The double-stage method shows larger prediction errors for all sets of initial conditions. C. Environment: Polyvinyl Chloride Phantom Previously, we identied the dynamics of two different contact environments: the elastic rubber ball with Kc and Bc of relatively the same order ( 2000), and the sponge with a Bc that was about ten times larger than its Kc value. For both objects, the value of n was estimated to be approximately 1. In this section, we study the distinct dynamic behavior of a PVC phantom, which is characterized by an n value close to 2. Fig. 4 illustrates the PVC phantom used for experiments. PVC phantoms have been used to mimic tissue properties for various applications, including experimental analysis and evaluation of surgical needle insertion methods [21]. The stiffness of the PVC phantom can be adjusted by changing the ratio of plastic to softener (or hardener). The PVC phantom used for this experiment

was constructed of ve portions of plastic and two portions of softener.4 For more information on the relationship between the stiffness of the PVC and the proportion of the plastic and the softener or hardener, see [22]. In [2], Yamamoto et al. used KV and HC methods, as well as polynomials of orders two, three, and four, and a secondorder polynomial plus viscous friction to model the dynamic characteristics of a PVC tissue phantom. The results in [2] have shown that the second-order polynomial plus viscous friction predicts the contact force with the same accuracy as that of the HC model and is more accurate than the KV and other polynomial models. Therefore, in addition to the KV and HC models, we also report the results of dynamic parameter estimation of this model, which we call extended KelvinVoigt (EKV) model, mathematically expressed by F (t) = (t), Ke 1 x(t) + Ke 2 x2 (t) + Be x 0, x(t) 0 x(t) < 0. (14)

For identication, we used the same EWRLS estimation method that has been used for KV models. 1) Initial Conditions: Following the strategy explained in Section V-A, three sets of initial values for the HC model parameters, two sets of initial values for the KV model parameters, and two sets of initial conditions for the EKV model parameters are considered for the estimation of the PVC phantom dynamic parameters as listed in Table V. 2) Validity of the Single-Stage Method: Fig. 14 shows the HC estimated parameters for the PVC phantom using the singlestage parameter estimation method in three repetitions of the experiment with the rst set of the initial conditions (set I). Although there are differences between the estimated values in
4 Supplier:

M-F Manufacturing Co., Inc., Fort Worth, TX.

HADDADI AND HASHTRUDI-ZAAD: REAL-TIME IDENTIFICATION OF HUNTCROSSLEY DYNAMIC MODELS OF CONTACT ENVIRONMENTS

563

TABLE V PVC PHANTOM: PARAMETER INITIAL CONDITIONS FOR THREE ESTIMATION MODELS

Fig. 14. PVC Phantom. Estimated parameters for the HC model using the single-stage method for three repetitions of real-time experiment with the rst set of initial conditions.

eters, the single-stage method for the HC model and the KV model are robust to initial conditions in terms of consistency. c identied using the single-stage c and B The parameters K method converge to small values for all initial condition sets in the rst few milliseconds after contact. Then, all the parameters converge to a constant higher value. However, the double-stage method results in different estimated parameters for the selected large range of initial conditions. The KV model displays a larger convergence rate compared with the single-stage method. The EKV model estimation results in the fastest convergence rate for Ke 1 and Be . However, the parameter Ke 2 does not converge to any consistent value for different initial conditions. This is primarily due to small penetration x and, as a result, small x2 which naturally decreases the contribution of its coefcient Ke 2 in the model-predicted force. The HC model, on the other hand, better approximates the dynamic behavior of the PVC tissue with lower %RMSE compared with those of the KV and EKV models, as reported in Table VI. VI. EFFECT OF EXCITATION LEVEL OF THE PROBE TRAJECTORY In this section, the effect of the level of excitation of the position command on the identication of the elastic rubber ball dynamics is studied. In previous sections, a combination of three different sinusoidal signals has been used to assure the richness of the probe trajectory. In this section, we study the estimation convergence when the desired trajectory consists of one or a combination of two sine signals. As discussed in Section III-B, since the model has three parameters to identify, as a rule of thumb, a combination of two sine waves would be sufcient [20]. However, due to the nonlinearity effect of the logarithm function, we expect to achieve accurate identication results for one sine signal as well. This effect will be explored next. A. Identication With Trajectory Commands Consisting of Two and One Sinusoidal Components The trajectory commands that are considered for the probe are x0 + 1.8 sin(11t) + 1.8 sin(15t) and x0 + 1.8 sin(11t), where the bias x0 ensures that the probe smoothly penetrates 1 cm into the material. The contact is the rubber ball, and the initial conditions are identical to those used in Section V-A. Fig. 17 shows the identication results for the HC model using the single-stage method. As expected for the HC model, the trajectory with two sine waves is rich enough and the parameters converge to similar values for the case where three sinusoidal waves were used, as reported in Fig. 8. Using a single sine component to build the trajectory command, the parameters still converge but at a substantially slower rate, i.e., after 500 s (the gure only shows the rst 120 s). The results from the single sine input are consistent with the results reported in Section V-A, while the excitation level of the trajectory command is not high enough for a general linear system with three parameters. This behavior can be related to the effect of nonlinearity of the ln() function in making the excitation richer. For the KV model, only two

Fig. 15. PVC Phantom. Convergence condition for the proposed single-stage 1K should method. For reliable estimated parameters, the condition x < 0 .B be met.

different repetitions of the experiment, we found that the results using the single-stage method are more consistent than the identied values using the double-stage method.5 One reason for the larger differences between the identied values of the PVC phantom for different trials, compared with those of the rubber ball and the sponge, is that it is more difcult for the robot to penetrate the PVC phantom; therefore, the excitation level of the identication input signal is affected accordingly. For PVC phantom, the value of n is close to 2, which is in contrast with its corresponding value n 1 for the rubber ball and the sponge. In order to conrm the validity of the single-stage method for the PVC environment, the approximation condition (10) is examined for one set of collected data and one set of inic and tial conditions. Fig. 15 suggests that the nal values for K c satisfy the convergence condition (10), and the estimation B is valid after approximately 300 ms from the contact moment t = 2 s. However, the condition is violated for a few samples at t = 8.7 s. 3) Effect of Initial Conditions and Comparing Hunt Crossley, KelvinVoigt, and Extended KelvinVoigt Models: In this section, the effect of initial conditions on the identication of the PVC phantom is examined. Fig. 16 shows the identied parameters and the prediction error for the HC models identied with the single-stage and double-stage methods, as well as for the KV and EKV models. Table VI summarizes the %RMSE values of the force prediction error. As shown by the param5 Because of space considerations, the double-stage results for various repetitions are not reported.

564

IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO. 3, JUNE 2012

Fig. 16. PVC Phantom. Estimated parameters and the force prediction errors for various initial conditions for the HC, KV, and EKV models. The KV model has one less parameter to identify and display. TABLE VI PVC PHANTOM: RMSE% OF THE HC AND KV MODELS

Fig. 18. Rubber Ball. Estimated parameters for the KV model and the force prediction error for different initial conditions when the trajectory command consists of one and two sinusoidal components.

Fig. 17. Rubber Ball. Estimated parameters for the HC model and the force prediction error for different initial conditions using the single-stage method when the trajectory command consists of one and two sinusoidal components.

parameters need to be identied; even one sinusoidal component is rich enough for convergence, as shown in Fig. 18. In some real-time applications, such as in-vivo estimation of soft tissues with nonlinear viscoelastic properties, the convergence time for a more accurate HC model may be too long in the absence of rich trajectories (sufcient level of excitation). On the other hand, a linear KV model that does not exhibit the nonlinear viscoelastic characteristics of soft tissues might serve

as a better alternative, since fewer dynamic parameters need to be identied, the model parameters may converge faster. The decision to address the tradeoff between convergence and accuracy in such situations depends on the dynamic nonlinearity of the material, the speed and richness of the trajectory, and the application at hand. VII. EFFECT OF PARAMETER VARIATIONS In many applications, robotic tasks include contact with different objects or with a nonhomogenous object which can exhibit varying dynamic properties of contact across different parts of the object. An example of such a task is robot-assisted softtissue palpation, which is used to detect a tumor in an organ or

HADDADI AND HASHTRUDI-ZAAD: REAL-TIME IDENTIFICATION OF HUNTCROSSLEY DYNAMIC MODELS OF CONTACT ENVIRONMENTS

565

estimation method to track variable dynamic parameters. This is because the covariance matrix P and the correcting gain become very small once the parameters converge. Therefore, after the rst convergence, the identied parameters cannot iteratively switch between the dynamic parameters of the two PVC objects as the process moves from one zone to the other. In order to solve the convergence issue, a self-perturbing (SP) least squares identication method [11], [23] is utilized in the single-stage method, which makes use of the following update rule for the covariance matrix: Pk +1 = 1 2 [Pk Lk +1 T k +1 Pk + NINT(ek 1 )I]

Fig. 19. Estimated parameters and force prediction error for intermittent contact (regular PVC and PVC with a hard inclusion) using (a) the regular singlestage method and (b) the SP version. (c) Penetration prole.

a calcied artery in the heart tissue [2]. In this section, we study the ability of the single-stage parameter estimation method to detect parameter variations when the robot is in intermittent contact between soft and hard tissues. To this purpose, we conduct an experiment to evaluate the performance of the proposed method during intermittent contact between two objects: 1) a regular PVC phantom representing a normal and healthy tissue and 2) a PVC phantom with a hard inclusion, i.e., an aluminum bar, to mimic (and exaggerate) the properties of organs with a tumor or tissue samples with calcied artery walls [2]. The 1-mm-thick aluminum bar inclusion is located at the depth of 7 mm from the surface of a 2-cm-thick PVC phantom. The prodding end-effector of the robot, serving as the palpation tool, penetrates inside the rst contact material, i.e., the PVC tissue phantom, at t = 2 s. After 5 s, the robot moves laterally for a few centimeters without sliding off the contact to face the second material, i.e., the PVC phantom with a hard inclusion. Intermittent exposures with these two materials are repeated every 5 s, and the parameters are identied in real time. The penetration prole in this intermittent experiment is shown in Fig. 19(c), where zone 1 represents contact with the regular PVC and zone 2 represents contact with the PVC with a hard inclusion. When the robot comes into contact with the regular PVC phantom at t = 2 s, the force prediction error increases as high as 2 N (shown up to 1 N) for 2 ms and then reaches 0.05 afterward within zone 1. In other instances, the robot is either in contact with the regular phantom or the phantom with the hard inclusion. The amount of force prediction error reaches an average value of 0.5 N and remains at this level after the contact changes to zone 2 as shown in Fig. 19(a). The high level of force prediction error and the constant nature of the estimated parameters after t = 10 s show that the EWRLS is not a suitable

where ek 1 := Fk T k k 1 , and I is an identity matrix of the same size as the matrix P. Therefore, whenever an abrupt change occurs and error increases, the adaptation gain increases automatically. Here, NINT is a roundoff operator, and and are design constants that can be adjusted according to the system measurement noise. The function NINT() grounds the SP term when e2 k 1 < 0.5, thereby implementing a regular EWRLS algorithm. Therefore, parameter controls the minimum error band for the SP action. In effect, determines a dead zone for noise rejection, while accepting sudden changes in the force prole. The value of is very much dependent on the characteristics of the two environments. Generally, parameter depends on the minimum difference between the two force levels that can be classied as two different objects. The value of parameter is dependent on the difference between the level of parameter changes, i.e., the greater the difference in parameter values, the more perturbation is required in the P matrix, and thus, the greater is the parameter . For our experiments, we choose = 0.05 and = 2.5 to achieve a balance between the noise or estimation error and the parameter variations. The parameter is an exponential forgetting factor that we reset to its initial 0 = 0.99, whenever NINT(e2 k 1 ) > 0, and allow it to exponentially reach a value of one to improve the speed of convergence. Using the SPRLS method, the average force prediction error decreases from 0.5 N in the rst 2 ms after contact to 0.05 N, as shown in Fig. 19(b). The estimated parameters are also dynamic and converge to levels that correspond to the material which the robot is in contact with. The average levels of the estimated parameters for the two contact materials are illustrated by dashed lines in Fig. 19(b). In the palpation scenario discussed in this section, it was assumed that the palpation device is always in contact with one of the two objects, either the PVC tissue phantom or the PVC tissue phantom with a hard inclusion. In the case when the intermittent transition takes place between a free motion regime and an object, the estimation process can be stopped once the robot leaves contact and resumed once the contact is restored. A reasonable solution to halt the identication process at loss of contact is by detecting the transition using force measurements and identifying when the contact force drops to zero. However, in practical cases, it would be difcult to identify exactly when the contact is lost due to the limitations in force measurements,

566

IEEE TRANSACTIONS ON ROBOTICS, VOL. 28, NO. 3, JUNE 2012

inaccurate calibration, or the presence of noise. As a result, small amount of force might still be detected during no-contact regime. A more robust method is to identify a suitable force threshold for the particular robot-environment transition, below which the contact is considered to be lost. VIII. CONCLUSION In this paper, we have rigourously evaluated a previously proposed single-stage parameter estimation method for the nonlinear HC model for different types of environments with unique dynamic characteristics under various experimental conditions. We, further, compared the results with those obtained from the double-stage method for the HC model and the EWRLS method for the KV model and its variant. For tissue phantom characterization, we also evaluated the proposed method against an EKV model. The experiments revealed that the single-stage method is more robust to parameter initial values and provides a superior identication performance in terms of force prediction error, compared with the double-stage method for the HC model and the EWRLS method applied to the KV and EKV models. In order for the single-stage method to be valid and consistent, the speed of operation of the robot must be changed according to the contact material. For instance, if the contact is highly damped, such as when the robot is in contact with a sponge, slower speed of operation is required; otherwise, the conditions that are required to ensure the consistency of the method will not be satised. This may limit the applicability of the method when there is limited control over the speed of operation or when high speed of operation is required. We also mimicked a robot-assisted palpation examination, where the robot intermittently established contact with a regular PVC phantom and a PVC phantom with a hard inclusion. To make the single-stage method responsive, an SP version of the RLS was implemented, which resulted in substantially smaller force prediction error and faster convergence of the estimation. REFERENCES
[1] H. Son, T. Bhattacharjee, and D. Lee, Estimation of environmental force for the haptic interface of robotic surgery, Int. J. Med. Robot. Comput. Assist. Surg., vol. 6, no. 2, pp. 221230, 2010. [2] T. Yamamoto, B. Vagvolgyi, K. Balaji, L. Whitcomb, and A. Okamura, Tissue property estimation and graphical display for teleoperated robotassisted surgery, in Proc. IEEE Int. Conf. Robot. Autom., 2009, pp. 4239 4245. [3] R. Ozawa, S. Arimoto, S. Nakamura, and J.-H. Bae, Control of an object with parallel surfaces by a pair of nger robots without object sensing, IEEE Trans. Robot., vol. 21, no. 5, pp. 965967, Oct. 2005. [4] S. E. Salcudean, K. Hashtrudi-Zaad, S. Tafazoli, S. DiMaio, and C. Reboulet, Bilateral matched impedance teleoperation with application to excavator control, IEEE Control Syst. Mag., vol. 19, no. 6, pp. 355367, Dec. 1999. [5] X. Mu and Q. Wu, On impact dynamics and contact events for biped robots via impact effect, IEEE Trans. Syst., Man, Cybern. B, vol. 36, no. 6, pp. 13641372, Dec. 2006. [6] K. Hashtrudi-Zaad and S. E. Salcudean, Adaptive transparent impedance reecting teleoperation, in Proc. IEEE Int. Conf. Robot. Autom., Minneapolis, MN, Apr.1996, pp. 13691373. [7] L. J. Love and W. J. Book, Force reecting teleoperation with adaptive impedance control, IEEE Trans. Syst., Man Cybern. B, vol. 34, no. 1, pp. 159165, Feb. 2004. [8] S. Singh and D. Popa, An analysis of some fundamental problems in adaptive control of force and impedance behavior: Theory and experiments, IEEE Trans. Robot. Autom., vol. 11, no. 6, pp. 912921, Dec. 1995.

[9] S. Misra and A. Okamura, Environment parameter estimation during bilateral telemanipulation, in Proc. IEEE Hapt. Symp., Washington, DC, 2006, pp. 301307. [10] D. Erickson, M. Weber, and I. Sharf, Contact stiffness and damping estimation for robotic systems, Int. J. Robot. Res., vol. 22, no. 1, pp. 41 57, Jan. 2002. [11] A. Haddadi and K. Hashtrudi-Zaad, Online contact impedance identication for robotic systems, in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Nice, France, 2008, pp. 974980. [12] N. Diolaiti, C. Melchiorri, and S. Stramigioli, Contact impedance estimation for robotic systems, IEEE Trans. Robot., vol. 21, no. 5, pp. 925935, Oct. 2005. [13] K. Hunt and F. Crossley, Coefcient of restitution interpreted as damping in vibroimpact, ASME J. Appl. Mech., vol. 42, pp. 440445, Jun. 1975. [14] D. Marhefka and D. Orin, A compliant contact model with nonlinear damping for simulation of robotic systems, IEEE Trans. Syst., Man Cybern. A, vol. 29, no. 6, pp. 566572, Nov. 1999. [15] A. Haddadi and K. Hashtrudi-Zaad, A new method for online parameter estimation of HuntCrossley environment dynamic models, in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2008, pp. 981986. [16] A. Achhammer, C. Weber, A. Peer, and M. Buss, Improvement of modelmediated teleoperation using a new hybrid environment estimation technique, in Proc. IEEE Int. Conf. Robot. Autom., 2010, pp. 53585363. [17] S. Jeon and S. Choi, Stiffness modulation for haptic augmented reality: Extension to 3D interaction, in Proc. IEEE Hapt. Symp., 2010, pp. 273 280. [18] G. Gilardi and I. Sharf, Literature survey of contact dynamics modelling, Mech. Mach. Theory, vol. 37, no. 10, pp. 12131239, 2002. [19] G. Goodwin and K. Sin, Adaptive Filtering, Prediction, and Control. Englewood Cliffs, NJ: PrenticeHall, 1984. [20] J.-J. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs, NJ: PrenticeHall, 1989. [21] S. P. DiMaio and S. E. Salcudean, Needle steering and motion planning in soft tissues, IEEE Trans. Biomed. Eng., vol. 52, no. 6, pp. 965974, Jun. 2005. [22] A. Baghani, H. Eskandari, S. Salcudean, and R. Rohling, Measurement of viscoelastic properties of tissue-mimicking material using longitudinal wave excitation, IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 56, no. 7, pp. 14051418, Jul. 2009. [23] D. J. Park and B. E. Jun, Self-perturbing recursive least squares algorithm with fast tracking capability, Electron. Lett., vol. 28, no. 65, pp. 558559, Mar. 1992. Amir Haddadi (S08) received the Ph.D. degree in electrical and computer engineering from Queens University, Kingston, ON, Canada, in 2011. He is currently a Postdoctoral Research Fellow with the University of British Columbia, Vancouver, BC, Canada. His current research interests include humanrobot interaction; control system design for teleoperation systems and haptic interfaces with biomedical applications; adaptive, robust, and nonlinear control system design; and system dynamics identication. Keyvan Hashtrudi-Zaad (SM08) received the B.Sc. degree from Sharif University of Technology, Tehran, Iran, the M.A.Sc. degree from Concordia University, Montreal, QC, Canada, and the Ph.D. degree from the University of British Columbia, Vancouver, BC, Canada, all in electrical engineering. He held a consulting position with Motion Metrics International Corp., Vancouver, on the development of a dynamic payload monitoring system for heavy-duty hydraulic machines. He then joined the Department of Electrical and Computer Engineering, Queens University, Kingston, ON, Canada, in 2001, where he is now an Associate Professor. His research interests include control systems design for telerobots and haptic interfaces, medical applications of haptics and virtual environments, system dynamics modeling, and biomedical robotics. He is currently an Associate Editor of the IEEE TRANSACTIONS ON HAPTICS. Dr. Hashtrudi-Zaad has served as the Technical Program Chair of the Robotics and Control Symposium of the IEEE Canadian Conference on Electrical and Computer Engineering, the Invited Session Chair of the IEEE Conference on Control Applications, and the Tutorial and Workshop Chair of the IEEE International Conference on Mechatronics and Automation.