Complexity

Volume 2019, Article ID 1406534, 16 pages

https://doi.org/10.1155/2019/1406534

## Position/Force Tracking Impedance Control for Robotic Systems with Uncertainties Based on Adaptive Jacobian and Neural Network

School of Electrical Engineering, Zhengzhou University, Zhengzhou 450001, China

Correspondence should be addressed to Tianlei Ma; nc.ude.uzz@amlt

Received 7 September 2018; Revised 27 November 2018; Accepted 20 December 2018; Published 3 January 2019

Academic Editor: Marcin Mrugalski

Copyright © 2019 Jinzhu Peng et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

In this paper, an adaptive Jacobian and neural network based position/force tracking impedance control scheme is proposed for controlling robotic systems with uncertainties and external disturbances. To achieve precise force control performance indirectly by using the position tracking, the control scheme is divided into two parts: the outer-loop force impedance control and the inner-loop position tracking control. In the outer-loop, an improved impedance controller, which combines the traditional impedance relationship with the PID-like scheme, is designed to eliminate the force tracking error quickly and to reduce the force overshoot effectively. In this way, the satisfied force tracking performance can be achieved when the manipulator contacts with environment. In the inner-loop, an adaptive Jacobian method is proposed to estimate the velocities and interaction torques of the end-effector due to the system kinematical uncertainties, and the system dynamical uncertainties and the uncertain term of adaptive Jacobian are compensated by an adaptive radial basis function neural network (RBFNN). Then, a robust term is designed to compensate the external disturbances and the approximation errors of RBFNN. In this way, the command position trajectories generated from the outer-loop force impedance controller can be then tracked so that the contact force tracking performance can be achieved indirectly in the forced direction. Based on the Lyapunov stability theorem, it is proved that all the signals in closed-loop system are bounded and the position and velocity errors are asymptotic convergence to zero. Finally, the validity of the control scheme is shown by computer simulation on a two-link robotic manipulator.

#### 1. Introduction

The development of technology has aroused people’s ever-growing interest on the safer and reliable, higher accuracy of robots. Many position tracking control methods have made some achievements in various application fields in recent years, such as cutting, grinding, welding, and so on [1–3]; the insufficient compliance for manipulator is always a key problem and has been given a lot of attention in robotics field, especially in complex and accurate tasks and even in collaboration with humans [4, 5]. Therefore, the applications of human-robot interaction (HRI) have become a new development domain and direction [6, 7]. It is necessary to develop an interaction control method that achieves position tracking and reliably adapts the force exerted on the environment in order to avoid damage both in the environment and in the manipulator itself [8].

The traditional force control methods proposed since 1980s can be mainly classified into two categories, namely, hybrid position/force control [9] and impedance control [10]. Impedance control method is to apply the motion trajectory and contact force into one dynamic framework, which can avoid the separate control processes of positions and force. In general, compared with the hybrid position/force control, the impedance control method can reduce the complexity of controller design and has better adaptability and robustness in the complex dynamic tasks [11]. Hence, in order to achieve the stable force tracking control, Chan and Yao [12] integrated sliding mode control into the impedance control method, which includes ideal impedance relationship in the sliding mode. Seul et al. [13, 14] proposed a series of force tracking impedance control methods. In [13] the impedance control schemes were classified into torque-based and position-based impedance methods based on different implementations of impedance function. Then, the manipulator was controlled in the free space and in the contact space [14], where the impedance function was improved to achieve position and force tracking, and the force error could be converged to zero for any environmental stiffness by using an adaptive technique. The force tracking performance of these methods, however, greatly depends on the accurate environmental information and the precise robotic system model.

Recently, many researchers tried to introduce the intelligent control methods into the force control to improve the tracking performance and robustness of the robotic system [15–20]. In [21], the neural network control technique was applied in impedance controller to compensate the uncertainties in an online manner. Li and Liu [22] designed an adaptive impedance hybrid controller, which could implement the desired contact force and track the command position in orthogonal subspace without precise environmental information. Jhan and Lee [23] proposed an adaptive fuzzy NN-based impedance control, where the adaptive fuzzy NN was used to approximate the robot dynamical model, so that the actual parameters of the manipulator need not be precisely known. In [24], a fuzzy logic system was applied as an approximator to estimate the unknown system dynamics, and then the proposed adaptive fuzzy back-stepping position/force control method could ensure all the signals of the close-loop system ultimately uniformly bounded. Duan et al. [25] presented an adaptive variable impedance control, which can adapt the environmental stiffness uncertainties. However, the contact force overshoot, which generated from the contact between robot and environment, was usually unconcerned in the above methods design. An improved impedance relationship was proposed in [26, 27] to reduce the contact force overshoot and to achieve the direct accurate time-varying force tracking. Roveda et al. [28, 29] calculated the parameters online in external controller to reduce the contact force overshoot. The methods can reduce the contact force overshoot provided that the model parameters of robotic systems are precisely known.

Since the model parameters of robotic systems can not be precisely obtained generally, lots of methods such as the adaptive Jacobian scheme, neural networks, back-stepping technology were employed to facilitate the tracking control of the manipulator with kinematic and dynamic uncertainties [30–32]. Liang et al. [33] designed a task-space observer to estimate the task-space positions and velocities simultaneously while NNs were employed to further improve the control performance through approximating the modified robot dynamics. These methods designed in Cartesian space can avoid inverse kinematical solution; however, there were only the position control problems included. To obtain the force control in Cartesian space, Jung et al. [13] proposed the position-based impedance control method, which can introduce the position-based control methods into the impedance control schemes. Literatures [27, 34, 35] used two-loop controllers, which included the inner-loop position control and the outer-loop force control, to achieve the position and force tracking control. Based on these methods, Bonilla et al. [36] proposed an inverse dynamical control law based on impedance control to achieve the position path tracking in both free and constrained spaces, which mainly focused on developing the compliant control scheme for constrained path tracking; however, the force tracking was not involved.

Consider the above drawbacks, in this paper, an adaptive neural network position/force tracking impedance controller is proposed for controlling robotic system with uncertainties in both free and contact spaces, where the force tracking can be achieved by position-based control method and the force overshoot can also be reduced efficiently. Consequently, the main contributions of this paper are presented as follows:

(1) Two-loop control architecture is presented to facilitate the position/force tracking impedance control for robotic systems, which are the outer-loop force impedance control and the inner-loop position tracking control.

(2) In the outer-loop, the improved impedance relationship based on PID-like scheme is proposed to reduce force overshoot when the end-effector contacts with environment.

(3) In the inner-loop, the adaptive Jacobian and RBFNN methods are employed to compensate the system model parameter and dynamical uncertainties. Then, the approximation errors of RBFNN and the external disturbances are restrained by a robust term. In this way, the desired contact force and the desired position trajectories can be tracked efficiently by the proposed intelligent position-based impedance control method.

(4) Based on the Lyapunov stability theorem, it is proved that all the signals in the closed-loop system are bounded, the position and velocity errors can be asymptotically converged to zero, and the contact force can be tracked to the desired force.

The reminder of this paper is organized as follows. In Section 2, some theoretical preliminaries are addressed, which consist of the mathematical notations, the RBFNN, and the details on dynamics of robotic system. In Section 3, the outer-loop position-based force tracking impedance control is designed to regulate interaction force when the robot contact with environment. Based on the reference trajectory generated from the outer-loop force impedance control, the adaptive neural network position tracking control scheme and its stability analysis are derived in Section 4. In Section 5, simulation results are presented to verify the effectiveness of the proposed method, and conclusions are drawn in Section 6.

#### 2. Problem Formulation and Preliminaries

In this paper, standard notations are used. We denote as the real number set, as the n-dimensional vector space, and represents the real matrix space. The norm of vector and that of matrix are defined as and , respectively. If is a scalar, let denote the absolute value. and are the minimum and the maximum eigenvalue of matrix , respectively. is the identity matrix.

##### 2.1. Description of RBFNN

In general, RBFNN has fast learning convergence speed and strong capability and has been proved in theory that RBFNN can approximate any nonlinear continuous function over a compact set to arbitrary accuracy [37]. The RBFNN structure is described as follows:where is the input vector, is the output vector, is the control input dimension, is the neuron node number, is the output dimension, is the weight matrix with , , is the RBFNN active function with hidden layer output function , and the Gaussian function is chosen as follows:where , is the center of the th neuron node, and is the width of the th neuron.

Numerous results indicate that for any continuous smooth function over a compact set , applying RBFNN (1) to approximate , if is sufficiently large, a set of ideal bounded weights exist, and we havewhere is the RBFNN reconstruction error. Since the ideal weight is unknown, the estimated weight is generally used to replace to approximate the unknown, continuous, nonlinear function; that is,where is the estimated weight matrix and can be trained by a weight learning law. We assume that is existed, and ideal parameter is in the compact set , which is defined as . The ideal parameter can be given as

*Assumption 1. *The RBFNN reconstruction error is bounded and satisfiedwhere is a positive constant.

In this paper, to ensure the control performance of manipulator, the RBFNN is used as a compensator to eliminate the dynamical uncertainties and the uncertain term of adaptive Jacobian in robotic system.

##### 2.2. Robotic Manipulator Dynamics and Properties

Consider a general -degree of freedom (DOF) rigid robotic manipulators with uncertain dynamics, the dynamical model can be described aswhere , , and represent the joint angular position vectors, velocity vectors, and acceleration vectors of the manipulator, respectively; is the positive definite and symmetric inertia matrix; represents the effect of centrifugal and Coriolis forces; is the gravity vector; is the friction effects; denotes the bounded unknown disturbances including unknown payload dynamics and unstructured dynamics; is the torque input vector; is the interaction torque vector when the manipulator contacts with environment; represents the Jacobian matrix from join space to task space; and is the contact force at the end-effector.

The following properties and assumption are required for the subsequent development.

*Property 2. *The inertia matrix is positive definite and symmetric, which is uniformly bounded and satisfies where and are some positive constants.

*Property 3. *The matrix is skew-symmetric; i.e.,

*Property 4. *The norm of the Coriolis and centrifugal forces matrix is bounded and satisfies where a positive constant.

*Assumption 5. *The unknown disturbance term is bounded by where is a positive constant.

Let be the position vector of the end-effector in task space. The relation between task space and joint space can be described by forward kinematics aswhere is the forward kinematics map, generally a nonlinear transformation between task space and joint space. The task-space velocities of end-effector is related to joint velocities as

*Assumption 6. *In general, a manipulator should work in a finite task space. The matrix is the inverse matrix of the Jacobian matrix when . When , the inverse matrix of the can be represented aswhere denotes the generalized inverse matrix of . Similar notations hold for the estimate Jacobian matrix as detailed later in Section 4.

In general, the uncertainties of the model parameters and the robot dynamics decreased the control performances of the robotic system directly. In this paper, an improved impedance relationship is designed to derive the reference trajectory planning scheme so that the reference position trajectories can be then generated. Then, an intelligent robust position-based impedance control scheme is proposed to achieve position trajectory tracking performance and the contact force tracking performance, where an adaptive Jacobian and RBFNN methods are designed to compensate the system uncertainties of robotic manipulator.

#### 3. Design of Outer-Loop Force Tracking Control Based on Improved Impedance Relationship

Impedance control method regulates the relationship between the position and force by selecting suitable impedance parameters. According to literature [10], the traditional impedance relationship of the robotic system satisfieswhere is the planned reference trajectory of end-effector for position control, which determined from environmental position and parameters, impedance parameters, and desired contact force; , , and are the desired inertia, damping, and stiffness matrices, respectively; is the force exerted on the environment at the end-effector.

In general, the contact force is determined by environmental stiffness and environmental damping; therefore, a second-order nonlinear function is used to approximate the environmental model, which can be expressed aswhere and denote the diagonal symmetric positive definite environmental stiffness and damping matrices, respectively, and is the environment position vector. Assume that the environment position is a constant, we have , then (16) can be rewritten as

##### 3.1. Reference Trajectory Planning

The force tracking response cannot generally be achieved quickly by using the traditional impedance control method shown in (15); when the end-effector of manipulator contacts with environment, the force overshoot may result in task failures. Therefore, in this paper, a PID-like impedance relationship is designed to improve the force tracking performance, which can be expressed aswhere , , and are the diagonal symmetric positive definite parameter matrices. The introduced PID-like force compensation can achieve a better expectation than the pure force in (15) so that the contact force generated at the end-effector can quickly converge to the desired value and reduce the force overshoot.

For convenience, we consider the force is exerted on one direction only. Replacing , , , , , , , , , , , , by , , , , , , , , , , , , , respectively, then the improved impedance function (18) can be rewritten asand the environment model (17) becomes

Define ; (19) and (20) can be rewritten asand

Taking Laplace transform to (21) and (22) yieldsand

Then, (24) can be rewritten as

Substituting (25) into (23) yields

Then, the force tracking in frequency domain can be obtained aswhereThen, steady state force tracking error can be obtained as

To ensure the steady state force error to be zero as the system approaches the stable equilibrium state, the reference position trajectory can be designed as

According to (30), it is obvious that the planned reference position trajectory is a dynamic function including the desired force , the environmental position , the environmental stiffness and damping , the impedance parameters , , , and the PID-like parameters , , . Assumed that the parameters in the improved impedance relationship (18) have been selected properly and the environmental information is accurately obtained, the reference position trajectory can be generated by the desired input force .

##### 3.2. Position-Based Impedance Control Scheme

In general, the traditional impedance control was classified into two methods: the position-based and the torque-based impedance control [13]. In the position-based impedance control, the outer-loop force impedance, and the inner-loop position tracking can be designed separately, where the force tracking performance mainly depends on the accuracy of the position tracking control in the inner-loop. Therefore, the position-based impedance control method has been widely applied in complex industrial systems such as servo control.

Denote the position command as the control input of the inner-loop position tracking control,where . Based on the improved impedance equation (18), the desired relationship between and can be represented as

Taking Laplace transform to (32) yields

To achieve the force tracking performance, the contact force is regulated to track the desired force ; an intelligent-based robust position tracking controller will be designed in Section 4.

*Remark 7. *The improved impedance relationship combines the PID-like method with the traditional impedance relationship in outer-loop force control to improve the response speed and the performance of force tracking. By choosing the appropriate PID parameters, the force tracking errors and the force overshoot can be reduced effectively with fast convergence when the manipulator contacts with environment. In addition, the improved impedance method can be applied to track a constant force or a twice-differentiable time-varying force.

*Remark 8. *Assumed that the proposed inner-loop position tracking controller is “perfect”, the manipulator can work well in the free and contact spaces based on the position-based impedance control scheme [27, 34, 35]. It means that the position command satisfies (31) if the forced direction is considered only. In this way, the position tracking and the force tracking can be achieved based on the two-loop separation design method in the position-based impedance control scheme.

#### 4. Design of Inner-Loop Position Tracking Control and Stability Analysis

In this section, an adaptive position tracking controller is proposed as the inner-loop in control system to track the command position trajectory generated from the outer-loop force impedance, where an adaptive Jacobian method is employed to approximate the task-space end-effector velocities and the interaction torques, and an adaptive RBFNN is designed to compensate the dynamic uncertainties of robotic systems and the uncertain Jacobian term. Based on Lyapunov theorem, the stability of the closed-loop robotic control system is then guaranteed.

##### 4.1. Adaptive Jacobian and RBFNN Position Tracking Controller Design

Define as the parameter vector in Jacobian matrix ; the task-space velocity of the end-effector and the robot interaction torque can be expressed aswhere and denote the velocity regressor matrix and the interaction torque regressor matrix, respectively. Note that the robot kinematic parameters uncertainties are always existed such as link length and mass so that the Jacobian matrix cannot be known precisely. Therefore, define the estimated Jacobian matrix , the estimation of and the estimation of can be represented aswhere denotes the estimated parameters vector. Then, the estimated task-space velocity error and estimated interaction torque error can be expressed aswhere .

Define a vector aswhere is a positive matrix and is the position tracking error of the end-effector. Then, differentiating (40) with respect to time yieldswhere is the velocity tracking error of the end-effector.

Define a filtered tracking error aswhere is the estimated value of the velocity tracking error .

Then, according to (36) and (40), we have

Differentiating (43), we obtainwhere and denote the derivative of and , respectively.

Next, define a filtered tracking error in joint space as

Assumed that the manipulator works in a finite task space, substituting (43) into (45) yieldswhere is the inverse of the estimated Jacobian matrix , which is similarly defined as in Assumption 5. Let a virtual joint velocity be

Substituting (47) into (46), we obtainand

Then, taking the derivative of (47) with respect to time, we obtain

Considering the uncertain , (41) can be redefined as

Then, (50) can also be redefined as

Substituting (50) into (52), according to (34), (36), (41), and (51), yields

Substituting (53) into (49) yields

Define that the states are and and then multiply with (54) on both sides; we havewhere , and

*Remark 9. *In general, the precise values of the robotic matrices , , , and are difficult to acquire but bounded [38]. Moreover, the Jacobian matrix is also bounded, it can be concluded that is bounded according to Assumption 6. Therefore, the unknown nonlinear function in (56) is bounded and can be approximated by using RBFNN.

In this paper, the unknown function is approximated by using RBFNN,where denotes the activation function of RBFNN, represents the ideal weight matrix, and denotes the minimum reconstructed error vector. Then, the adaptive RBFNN position trajectory tracking control law can be given aswhere and are the controller position and velocity gain matrices, respectively, and denotes the robust compensation term which is used to compensate the approximation error of RBFNN and external disturbances. Substituting the control law (58) into (55), the closed-loop error equation of the robotic system is as follows:

Sowhere denotes the weight estimated error.

Then, substituting (39) and (60) into (59) results inAccording to Assumptions 1 and 5, the RBFNN modeling errors are bounded as

##### 4.2. Stability Analysis

Theorem 10. *According to the robotic dynamic model (7), assume that Assumptions 1–6 are all satisfied, the adaptive RBFNN position tracking control law is designed as (58), where the robust compensation term can be given byBy the projection algorithm, the adaptive updating law for the Jacobian matrix parameter and the weight matrix of RBFNN are designed as follows:where and are both the positive matrices.**Then, the filter error , the task-space position error , the adaptive Jacobian matrix parameter error , and the adaptive RBFNN weight matrix error are all bounded and the contact force can converge to desired force as .*

*Proof. *Choose a Lyapunov function candidate asDifferentiating (66) with respect to time and substituting (61) yieldAccording to the bounded modeling errors (62) and robust term (63) and the RBFNN adaptive law (65), considering the fact , we haveandAccording to Property 3, substituting (68) and (69) into (67) yieldsSubstituting (38), (39), (42), and (45) into (70) yieldsAccording to the adaptive updating law (64) and considering the fact , we haveSubstituting (72) into (71) yieldsAccording to (66) and (73), it can be concluded that the error signals , , , and are all bounded. That means and are both bounded, and is also bounded using (45). From (39), is bounded since the contact force is bounded. Then, according to (40) and (42), it can be concluded that and are bounded as the position controller input and is bounded, and is also bounded according to (47) as the estimate Jacobian matrix is nonsingular. Therefore, is bounded according to (46) which implies that is bounded. Next, it can be concluded that is bounded and is also bounded from (41) as is bounded. And , are both bounded according to (51) and (52). In addition, according to Property 4, it can be concluded that is bounded from (61) so that is also bounded from (49), and then is bounded, which means is bounded as is bounded. Considering (44), is bounded, which means is also bounded.

Differentiating (73) with respect to time, we havewhere denotes the derivative of . Since the error signals , , , are all bounded, is uniformly continuous. According to** Barbalat’s lemma** [39], we can conclude and as . Since is bounded, we can obtain as . That is,Based on the position-based impedance control scheme proposed in Section 3, we can concludewhen the manipulator contacts with environment.

According to the above analysis, the block diagram of the whole closed-loop impedance control system is shown as Figure 1.