In this paper, a novel adaptive robust approach to modeling and control of a class of flexible-arm robots subject to actuators unmodeled dynamics is proposed. It is shown how real-time signals measured from a dynamical system can be utilized to improve the accuracy of the mathematical model of flexible robots. Given the elasticity of the robot’s arms, flexible manipulators have both passive and active degrees of freedom. A nonlinear robust controller is designed for the active degrees of freedom to enable the robot to follow desired trajectories in the presence of actuators unmodeled dynamics. Furthermore, it is shown that under some feasible conditions, another nonlinear robust controller is designed for the passive degrees of freedom. Moreover, to use the system response for model extraction, two auxiliary signals are proposed to provide sufficient information for improving the accuracy of the dynamics of the system numerically. Additionally, two adaptive laws are proposed in each case to update the two introduced auxiliary signals. As a result, the controller controls the passive degrees of freedom after the active degrees of freedom converge to their desired trajectories. Simultaneously, the information collected from the system to update the auxiliary signals enhances the model accuracy. In the end, simulation results are presented to verify the performance of the proposed controller.
The area of flexible-arm robots has attracted much attention during the last few decades [
There has been a great number of studies coping with controlling flexible-arm robots, many of which investigate both theoretical and experimental aspects in this field [
Infinite dimensionality is one of the most considerable challenges in modeling flexible robots. The existence of flexibility in the system results in dynamics governed by partial differential equations. Thus, techniques such as modal truncation are employed to express the dynamics by a set of ordinary differential equations [
On account of the elasticity of the arms, there are degrees of freedom on which no actuation acts (Passive degrees of freedom). Consequently, flexible-arm robots fall into the category of underactuated mechanical systems. That is, systems with a lower number of control inputs compared to the number of degrees of freedom. Designing controllers for underactuated systems is an open problem. Control designs for such systems are dynamics-dependent while dynamical characteristics of the system play an essential role in the development of a control strategy. As an example of the existing control methods for underactuated systems, Mahmut Reyhanoglu et al. presented a theoretical scheme for modeling and control of underactuated systems. As one of the requirements of their approach, the non-integrable acceleration relations are required to be satisfied [
Feedback linearization methods are wildly used for nonlinear systems. Nevertheless, As a result of the underactuated nature of flexible arms, the exact feedback linearization method [
In this paper, a novel adaptive robust nonlinear controller is designed for a class of multi-link flexible arms subject to uncertainties and unmodeled dynamics of the actuators. The contribution of the current research work is to utilize the system’s real-time responses to improve the accuracy of the available mathematical model. Additionally, despite uncertainties and underactuated nature of the system, the proposed controller is able to track desired trajectories asymptotically. That is, despite the presence of uncertainties and disturbance sources, the tracking error converges to zero. To do such, first, an adaptive robust controller is designed for the active degrees of freedom. Meanwhile, adaptive laws are proposed to estimate the system uncertainties as well as approximating actuators uncertain dynamics. Furthermore, another adaptive robust controller is designed to control the passive degrees of freedom where the stabilities of the overall closed-loop system, in both cases, in the presence of uncertainties are established. In addition, due to the flexibility of the arms, the extraction of the system dynamics is demanding and, for the most part, should be done analytically. In the present approach, two adaptive signals are introduced and synthesized such that the cumbersome analytical part of the model extraction can be done numerically.
In the current study, a multi-link flexible-arm robot is considered for modeling and control purposes. The investigated system is assumed to include m links containing l elastic modes. Therefore, the total degrees of freedom come to n = m + l .
Equations governing a flexible arm using the Euler-Bernoulli beam can be written as follows
( M 11 M 12 M 21 M 22 ) ( q ¨ 1 q ¨ 2 ) + C ( q , q ˙ ) + G ( q ) = ( 0 T ) (1)
where M ( q ) ∈ ℝ n × n = ( M 11 M 12 M 21 M 22 ) is the positive definite inertial matrix;
C ( q , q ˙ ) ∈ ℝ n includes Coriolis and centrifugal terms; G ( q ) ∈ ℝ n contains gravitational terms; q T = [ q 1 q 2 ] T is the generalized coordinates vector where q 1 ∈ ℝ l denotes the passive degrees of freedom, and q 2 ∈ ℝ m represents the active degrees of freedom. T ∈ ℝ m is the torque generated by the actuators acting on the active degrees of freedom. C ( q , q ˙ ) and G ( q ) can be calculated as [
C ( q , q ˙ ) = M ˙ ( q ) q ˙ − 1 2 ∂ ( q ˙ T M ( q ) q ˙ ) ∂ q (2)
G ( q ) = ∂ U ∂ q (3)
where U denotes potential energy of the system. Extracting the dynamic equations of flexible robot arms is time consuming and can be analytically complicated. Due to its analytical nature, there exists a high chance of computational errors as the equations are derived. The extraction of the dynamic equations can be divided into two phases: 1) Extracting the portions of the dynamics that are straightforward and demand less calculations such as kinetic and potential energy. It is to be noted that the inertia matrix M ( q ) can be arrived at directly from the rearrangement of the kinetic energy of the system. 2) Obtaining the portions of the dynamics which require lengthy analytical calculations such as G ( q ) and C ( q , q ˙ ) . It is worth noting that calculating G ( q ) and C ( q , q ˙ ) require calculus of variations which can be cumbersome for systems with a large degrees of freedom. The idea of this paper is to propose a way to procure G ( q ) and C ( q , q ˙ ) numerically using the system responses. For this purpose, (1) is rewritten as
( M 11 M 12 M 21 M 22 ) ( q ¨ 1 q ¨ 2 ) + ( h 1 ( q , q ˙ ) h 2 ( q , q ˙ ) ) = ( 0 T + d ( t ) ) (4)
where
h 1 ( q , q ˙ ) = M ˙ 11 ( q ) q ˙ 1 + M ˙ 12 ( q ) q ˙ 2 − 1 2 ∂ ( q ˙ T M ( q ) q ˙ ) ∂ q 1 + ∂ U ∂ q 1 (5)
h 2 ( q , q ˙ ) = M ˙ 21 ( q ) q ˙ 1 + M ˙ 22 ( q ) q ˙ 2 − 1 2 ∂ ( q ˙ T M ( q ) q ˙ ) ∂ q 2 + ∂ U ∂ q 2 (6)
and d ( t ) is an unknown vector added to the equations to account for uncertain dynamics of the actuators. It is also assumed that ‖ d ( t ) ‖ ≤ γ where γ is an unknown constant.
In order to complete the dynamic equations given in (4), h 1 ( q , q ˙ ) and h 2 ( q , q ˙ ) should be computed. As it follows from (5) and (6), h 1 ( q , q ˙ ) and h 2 ( q , q ˙ ) are expressed in terms of the matrix M, system states, and its potential energy. Therefore, one can use the measurement of the states along with the kinetic and potential energy of the system and use numerical computations to obtain h 1 and h 2 . Hence, the system dynamics is computed as it runs without going through analytical calculations. However, such dynamic equations are prone to numerical and measurement errors. Thus, a special measure should be taken to improve the accuracy of such computations. To compensate for such errors, two unknown auxiliary signals P and S are employed as follows
M 1 1 q ¨ 1 + M 12 q ¨ 2 + H 1 P = 0 (7)
M 2 1 q ¨ 1 + M 2 2 q ¨ 2 + H 2 S = T (8)
where P ∈ ℝ l and S ∈ ℝ m are two unknown auxiliary signals introduced to account for numerical and measurement errors; H 1 ( q , q ˙ ) ∈ ℝ l × l , H 2 ( q , q ˙ ) ∈ ℝ m × m are two diagonal matrices constructed as H 1 ( q , q ˙ ) = d i a g [ h 1 2 , h 1 2 , ⋯ , h 1 n ] and H 2 ( q , q ˙ ) = d i a g [ h 2 2 , h 2 2 , ⋯ , h 2 n ] ; h i j denotes the jth element of the vector h i where i = 1 , 2 . Given the physical and geometric properties of flexible arm robots, it is not restrictive to assume that both H 1 and H 2 , as defined above, have full ranks. Therefore, their column space can be utilized as a basis for the vectors h 1 and h 2 at any given time. Thus, for any exact value of h 1 and h 2 at any given time, there exist two unique vectors P and S such that h 1 = H 1 P and h 2 = H 2 S . It is clear that the vectors P and S change in values over time. Hence, their values need to be updated according to the system response. Therefore, an adaptive procedure is developed for each of these vectors along with designing a controller in Section 3. Consequently, as the system runs, its response is used to update the values of P and S.
In this section, an adaptive controller is designed for the considered system in the presence of uncertainties. Alongside the adaptive controller, two adaptive laws are derived for updating the values of P and S which are used to complete the dynamic equations. As it was shown in the previous section, (7) and (8) exhibit the dynamic equations governing an n-degree of freedom flexible-arm, with m active and l passive degrees of freedom. The first objective is to design a controller for the active degrees of freedom.
Employing (7), (8) can be rewritten as follows
M 21 ( M 11 − 1 ( − M 12 q ¨ 2 − H 1 P ) ) + M 22 q ¨ 2 + H 2 S = T + d (9)
Two matrices N and R are defined as R = M 21 M 11 − 1 and N = M 22 − R M 12 . Consequently, (9) comes to
N q ¨ 2 + H 2 S − R H 1 P = T + d (10)
As N is positive definite, (10) can be expressed as follows
q ¨ 2 = − H ¯ 2 S + H ¯ 1 P + U ¯ + N − 1 d (11)
where H ¯ 2 = N − 1 H 2 ; H ¯ 1 = N − 1 R H 1 , and U ¯ = N − 1 T . Tracking error and its dynamics can be defined as e = q 2 − q 2 d and e ˙ = q ˙ 2 − q ˙ 2 d , respectively; q 2 d is the desired trajectory for the active degrees of freedom. A measure of the tracking error is also defined as
Q = λ 1 e + e ˙ (12)
where λ 1 ∈ ℝ m × m is a positive definite matrix. The estimation errors of the uncertainty vectors are shown by P ˜ and S ˜ which are defined as S ˜ = S ¯ − S and P ˜ = P ¯ − P ; S ¯ ∈ ℝ m and P ¯ ∈ ℝ l are the estimations of the uncertainty vectors. Using the definitions of S ˜ and P ˜ , (11) is rewritten as follows
q ¨ 2 = − H ¯ 2 ( S ¯ − S ˜ ) + H ¯ 1 ( P ¯ − P ˜ ) + U ¯ + N − 1 d (13)
Considering (12), Q ˙ is calculated as follows
Q ˙ = λ 1 ( q ˙ 2 − q ˙ 2 d ) − H ¯ 2 ( S ¯ − S ˜ ) + H ¯ 1 ( P ¯ − P ˜ ) + U ¯ + N − 1 d − q ¨ 2 d (14)
To proceed with the control design, the following Lyapunov function is proposed
V = 1 2 ( Q T Q + P ˜ T P ˜ + S ˜ T S ˜ + γ ˜ 2 ) (15)
where γ ˜ = γ ^ − γ is the Euclidean norm of the estimation error of the uncertain dynamics of the actuators. Taking time-derivative of (15) gives
V ˙ = Q T ( λ 1 q ˙ 2 − λ 1 q ˙ 2 d − H ¯ 2 ( S ¯ − S ˜ ) + H ¯ 1 ( P ¯ − P ˜ ) + T ¯ + u + N − 1 d − q ¨ 2 d ) + P ˜ ˙ T P ˜ + S ˜ ˙ T S ˜ + γ ˜ γ ˙ (16)
where U ¯ is decomposed into U ¯ = T ¯ + u . Given the structure of (16), the following control and adaptive laws are proposed
T ¯ = − ψ Q − λ 1 ( q ˙ 2 − q ˙ 2 d ) + H ¯ 2 S ¯ − H ¯ 1 P ¯ + q ¨ 2 d (17)
P ˜ ˙ T − Q T H ¯ 1 = 0 (18)
S ˜ ˙ T + Q T H ¯ 2 = 0 (19)
where ψ ∈ ℝ m × m is a positive definite matrix; (17) is the proposed control law, and (18)-(19) are the adaptive laws to update the two auxiliary signals introduced in (7)-(8). Substituting (17)-(19) into (16) yields
V ˙ = − Q T ψ Q + Q T u + Q T N − 1 d + γ ˜ γ ˜ ˙ (20)
(20) can be rewritten as follows
V ˙ ≤ − λ m i n ( ψ ) ‖ Q ‖ 2 + ‖ Q ‖ ‖ N − 1 ‖ ( γ ^ − γ ˜ ) + Q T u + γ ˜ γ ˜ ˙ (21)
where ‖ . ‖ stands for the Euclidean norm. Since N is a bounded positive definite matrix, the adaptive law to update γ ˜ is proposed as
γ ˜ ˙ = ‖ Q ‖ ‖ N − 1 ‖ (22)
Employing (22), (21) becomes
V ˙ ≤ − λ m i n ( ψ ) ‖ Q ‖ 2 + ‖ Q ‖ ‖ N − 1 ‖ γ ^ + Q T u (23)
To establish the asymptotic stability of the closed-loop system, the second part of the control law is introduced as follows in (24). It is to be noted that the second part of the control law compensates for the actuators uncertain dynamics.
u = − γ ^ 2 ‖ N − 1 ‖ Q γ ^ ‖ Q ‖ ‖ N − 1 ‖ + σ ( t ) (24)
where σ ( t ) ∈ L 1 is an arbitrary signal where ∫ 0 ∞ σ ( t ) d t < ∞ . By substituting (24) into (23), the following is obtained
V ˙ ≤ − λ min ( ψ ) ‖ Q ‖ 2 + γ ^ ‖ Q ‖ ‖ N − 1 ‖ σ ( t ) γ ^ ‖ Q ‖ ‖ N − 1 ‖ + σ ( t ) (25)
Since ‖ Q ‖ , ‖ N − 1 ‖ , σ ( t ) are positive, the following inequality is concluded
γ ^ ‖ Q ‖ ‖ N − 1 ‖ σ ( t ) γ ^ ‖ Q ‖ ‖ N − 1 ‖ + σ ( t ) ≤ σ ( t ) (26)
Hence, (25) is simplified as follows
V ˙ ≤ − λ m i n ( ψ ) ‖ Q ‖ 2 + σ ( t ) (27)
Taking integral from both sides of (27) gives
λ m i n ( ψ ) ∫ 0 t ‖ Q ‖ 2 d τ ≤ V ( 0 ) − V ( t ) + ∫ 0 t σ ( τ ) d τ (28)
Considering the fact that σ ( t ) ∈ L 1 , the following is achieved
λ min ( λ 2 ) ∫ 0 ∞ ‖ Q ‖ 2 d t ≤ V ( 0 ) − V ( ∞ ) + a (29)
where a is a positive constant. It is observed that the left side of (29) is non-negative, thus, it follows that ( V ( 0 ) − V ( ∞ ) ) is bounded. As a result, it is deduced that the right side of (29) remains bounded as well. Consequently, Q ∈ L 2 . It is also easy to check that Q ˙ ∈ L ∞ . Therefore, it is concluded from Barbalat’s lemma that lim t → ∞ Q ( t ) = 0 which implies lim t → ∞ e ( t ) = 0 . Thus, it was proven that adopting the extracted control laws (17), (24), and adaptive laws (18), (19)) and (22) guarantees the asymptotic stability of the closed-loop system. That is, the system is able to track desired trajectories asymptotically in the presence of model and actuators uncertainties.
In this section, a robust nonlinear controller is designed for the passive degrees of freedom alongside extracting adaptive laws to update uncertainties vectors. To do such, (8) is rewritten as
q ¨ 2 = M 22 − 1 ( − M 21 q ¨ 1 − H 2 S + T + d ) (30)
Substituting (30) into (7) yields
F q ¨ 1 − G H 2 S + h 1 P + G T + G d = 0 (31)
where G = M 12 M 22 − 1 ; F = M 11 − M 12 M 22 − 1 M 21 . It follows from the structure of M that M22 is nonsingular. In addition, it can be easily shown that F is a positive definite matrix as well. Therefore (31) can be expressed as follows
q ¨ 1 = K S − D P − V − F − 1 G d (32)
where K = F − 1 G H 2 , D = F − 1 H 1 , J = F − 1 G T , and J is a control signal yet to be defined. The actual control input can be retrieved from J as follows
T = G − 1 F J (33)
It is concluded from (33) that G should be nonsingular. Thus, according to the definition of G, M12 needs to be invertible as well. Hence, in order to extract a non-singular control law for the passive degrees of freedom, the following assumption is made:
Assumption 1 It is assumed that the number of passive degrees of freedom is less or equal than the number of active degrees of freedom.
The tracking error for the passive degrees of freedom and its dynamics are defined as e ¯ = q 1 − q 1 d and e ¯ ˙ = q ˙ 1 − q ˙ 1 d where q 1 d is the desired trajectory for the passive degrees of freedom. Moreover, a measure of the tracking error is defined as Q ¯ = λ 2 e ¯ + e ¯ ˙ where λ 2 ∈ ℝ l × l is a positive definite matrix. Using the definitions of S ˜ , P ˜ , which are defined in the previous section, (32) is expressed as follows
q ¨ 1 = K ( S ¯ − S ˜ ) − D ( P ¯ − P ˜ ) − J − F − 1 G d (34)
Based on the structure of (34), the following Lyapunov function is introduced
V ¯ = 1 2 ( Q ¯ T Q ¯ + P ˜ T P ˜ + S ˜ T S ˜ + γ ˜ 2 ) (35)
The time-derivative of (35) is calculated as
V ¯ ˙ = Q ¯ T ( λ 2 ( q ˙ 1 − q ˙ 1 d ) + K S ¯ − D P ¯ − J 1 − J 2 ) + ( P ˜ ˙ T + Q ¯ T D ) P ˜ + ( S ˜ ˙ T − Q ¯ T K ) S ˜ + γ ˜ γ ˜ ˙ − Q ¯ T ( F − 1 G d − q ¨ 1 d ) (36)
where J = J 1 + J 2 . Given (36), the following control and adaptive laws are proposed
J 1 = ψ ¯ Q ¯ + λ 2 ( q ˙ 1 − q ˙ 1 d ) + K S ¯ − D P ¯ − q ¨ 1 d (37)
P ˜ ˙ T + Q ¯ T D = 0 (38)
S ˜ ˙ T − Q ¯ T K = 0 (39)
where ψ ¯ ∈ ℝ l × l is a positive definite matrix. Substituting (37)-(39) into (36) and taking the Euclidean norm form both sides lead to the following
V ¯ ˙ ≤ − λ m i n ( ψ ¯ ) ‖ Q ¯ ‖ 2 + ‖ Q ¯ ‖ ‖ F − 1 ‖ ‖ G ‖ ( γ ^ − γ ˜ ) − Q ¯ T J 2 + γ ˜ γ ˜ ˙ (40)
To make (40) negative semi-definite, the following control and adaptive laws are proposed
J 2 = γ ^ 2 Q ¯ ‖ F − 1 ‖ ‖ G ‖ γ ^ ‖ Q ¯ ‖ ‖ F − 1 ‖ ‖ G ‖ + σ ( t ) (41)
γ ˜ ˙ = ‖ Q ¯ ‖ ‖ F − 1 ‖ ‖ G ‖ (42)
where σ ( t ) is a positive arbitrary signal belonging to L 1 . Substituting (40)-(42) into (40) gives
V ˙ ≤ − λ m i n ( ψ ) ‖ Q ‖ 2 + σ ( t ) (43)
Similar to the stability analysis presented in the previous section, it is concluded from (43) that lim t → ∞ Q ¯ ( t ) = 0 , which implies that lim t → ∞ e ¯ = 0 . Hence, it is guaranteed that the tracking error converges to zero asymptotically in the presence of uncertainties.
In this section, a simulation is conducted to validate the analytical results. The simulated case study is a one-link flexible-arm robot with a payload at the tip of the arm. It is assumed that the system is subject to uncertainties and unmodeled dynamics in the actuators. The strategy is, first, to control the active degrees of freedom to their desired values. Then, the control law and adaptive laws switch to the non-collocated control to dampen the vibrations of the flexible arm. The physical properties of the investigated arm are as follows: Arm s Length = 1 m; Density = 7850 kg/m3; Beams section area = 2e−4 m2; Elasticity = 2.07e11; Load mass = 0.1 kg. The actual values uncertainty signals are S = 8.2 and P = 5.5 .
It is shown in
In
Finally, the required actuator torque to control the arm is shown in
It was shown both analytically, in Section 3, and through simulation that the proposed controller can achieve asymptotic tracking in the presence of unknown
and bounded disturbance sources. As a comparison to other recent robust approaches to the control problem of such flexible manipulators, authors in [
As presented in detail, an adaptive robust nonlinear controller was designed for a class of flexible-arm robots based on the Lyapunov stability theory. A new numerical technique was presented to facilitate the extraction of the governing equations of flexible-arm robots. It was shown how computational errors caused by numerical operations could be compensated and approximated as uncertainty signals. In the control design process, first, an adaptive controller was designed for the active degrees of freedom to enable the system to follow desired trajectories in the presence of uncertainties. Further, it was shown that under some feasible conditions, another adaptive robust controller could be designed for the passive degrees of freedom. Therefore, as the system was being controlled, its responses were utilized to improve the accuracy of the mathematical model. Thus, the presented method leads to improvements in the controller performance and increases tracking accuracy. Furthermore, due to a numerical approach to extracting the dynamics, the amount of analytical computations is reduced dramatically.
Bagheri, P., Behjat, L. and Sun, Q. (2018) An Adaptive Robust Approach to Modeling and Control of Flexible Arm Robots. World Journal of Engineering and Technology, 6, 282-295. https://doi.org/10.4236/wjet.2018.62017