Shih-Feng Chen Imin Kao Manufacturing and Automation Laboratory Department of Mechanical Engineering State University of New York at Stony Brook Stony Brook, New York 11794-2300, USA [email protected] Abstract Conservative Congruence Transformation for Joint and Cartesian Stiffness Matrices of Robotic Hands and Fingers 1. Introduction In this paper, we develop the theoretical work on the properties and mapping of stiffness matrices between joint and Cartesian spaces of robotic hands and fingers, and propose the conservative congruence transformation (CCT). In this paper, we show that the conventional formulation between the joint and Cartesian spaces, Kθ = JθT Kp Jθ , first derived by Salisbury in 1980, is only valid at the unloaded equilibrium configuration. Once the grasping configuration is deviated from its unloaded configuration (for example, by the application of an external force), the conservative congruence transformation should be used. Theoretical development and numerical simulation are presented. The conservative congruence transformation accounts for the change in geometry via the differential Jacobian (Hessian matrix) of the robot manipulators when an external force is applied. The effect is captured in an effective stiffness matrix, Kg , of the conservative congruence transformation. The results of this paper also indicate that the omission of the changes in Jacobian in the presence of external force would result in discrepancy of the work and lead to contradiction to the fundamental conservative properties of stiffness matrices. Through conservative congruence transformation, conservative and consistent physical properties of stiffness matrices can be preserved during mapping regardless of the usage of coordinate frames and the existence of external force. KEY WORDS—stiffness control, conservative congruence transformation, stiffness mapping, differential Jacobian The International Journal of Robotics Research Vol. 19, No. 9, September 2000, pp. 835-847, ©2000 Sage Publications, Inc. The conventional formulation for the mapping of stiffness matrices between the Cartesian and joint spaces, Kθ = JθT Kp Jθ , (1) was first derived by Salisbury (1980) and generally has been accepted and applied. However, we will show in this paper that eq. (1) is only valid when robotic manipulators are at their unloaded equilibrium configuration, i.e., without external force. In this paper, we present the theoretical derivations of the general relationship of congruence transformation to replace eq. (1), as well as results of numerical simulation. This paper builds on the previous results of properties of grasp stiffness matrix via congruence transformation to discuss the fundamental properties of a grasp stiffness matrix as applied in the analysis of grasping and dextrous manipulation in linear R3×3 spaces (Salisbury 1980; Cutkosky and Kao 1989; Pigoski, Griffis, and Duffy 1992; Griffis and Duffy 1993; Ciblak and Lipkin 1994; Howard, Žefran, and Kumar 1998; Kao, Cutkosky, and Johansson 1997; Chen and Kao 1998; Kao and Ngo 1999). From the analysis of conservative properties (Kao and Ngo 1999) and the simulation results of congruence transformation of stiffness matrices (Chen and Kao 1998), the conventional formulation in eq. (1), defining the relationship of stiffness matrices via the Jacobian matrix between the joint and Cartesian spaces, is found to be a nonconservative mapping. This is proven by the work discrepancy of stiffness control due to the configuration-dependent Jacobian matrix (Kao and Ngo 1999). That is, there will be discrepancy between the work done in the joint and Cartesian spaces using the stiffness control in eq. (1), except in the following special cases: (1) when the Jacobian and stiffness 835 Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 836 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / September 2000 matrix are uncoupled such as that of a gantry (or Cartesian) robot manipulator, (2) when the manipulator is maintained at its unloaded position at all times (i.e., no external load), and (3) special Kθ or Kp (see Kao and Ngo 1999). In this paper, we will present the conservative congruence transformation (CCT) as the correct and general relationship for congruence mapping with stiffness control in robotics. Theoretical derivation and numerical results are used to illustrate the CCT as the complete transformation. We will also show that the conventional formulation in equation (1) is not valid if external force is applied. The CCT accounts for changes in grasping geometry as well as changes in forces and torques due to stiffness matrices. The change in geometry under the presence of external load is captured by the matrix, Kg , in the CCT that includes the differential Jacobian (a Hessian–3D tensor) and external load. It will be shown that the differential Jacobian plays an important role in determining the conservative property of stiffness control. In addition, we will discuss the properties of stiffness matrices when mapping between coordinate basis. In 1980, Salisbury first derived a congruence transformation to describe the mapping between the Cartesian and the joint stiffness matrices in robot manipulation. Žefran and Kumar (1997) presented an asymmetric connection and used the congruence transformation to define an affine connection in a symmetric Cartesian stiffness matrix. Kao and Ngo (1999) defined two conservative criteria of a linear stiffness matrix, which does not involve rotational components, to investigate the conservative properties of the congruence transformation and the corresponding control strategies for conservative control using the conventional formulation. Furthermore, twodimensional planar robot simulation results of nonconservative properties via congruence transformation were provided by Chen and Kao (1998, 1999). 3. Theoretical Background In this section, we will briefly describe the criteria for conservative stiffness matrix and the conventional formulation as presented by Salisbury (1980). After that, the conservative congruence transformation will be derived and discussed. 2. Literature Survey Robot compliance and stiffness have been studied by many investigators. Dimentberg (1965) employed screw theory, introduced by Ball (1900), to describe the motion of a rigid body in a general spatial potential field to derive a symmetric stiffness matrix at an unloaded equilibrium configuration. Lončarić (1985) applied Lie algebra to investigate the properties of symmetric stiffness matrices. Recently, Pigoski, Griffis, and Duffy (1992) studied a 6×6 asymmetric stiffness matrix with three elastic springs coupling between two planar rigid bodies. In addition, Griffis and Duffy (1993) derived a 6 × 6 asymmetric Cartesian stiffness matrix in the study of the Stewart platform. Howard, Žefran, and Kumar (1998) used the differential geometry and properties of Lie groups to show that the 6 × 6 stiffness matrix is asymmetric in any conservative system subjected to a nonzero external load. In Mussa-Ivaldi, Hogan, and Bizzi (1985); Hogan (1990); and Kao, Cutkosky, and Johansson (1997), a 2D grasp stiffness matrix was partitioned into two components pertaining to the conservative gradient and nonconservative curl functions. The stiffness matrices for human grasping tasks were obtained via experiments and decomposed into symmetric (conservative) and skew symmetric (nonconservative) components in R2×2 (Kao, Cutkosky, and Johansson 1997). Ciblak and Lipkin (1994) developed a formulation to represent the skew symmetric part of the 6 × 6 Cartesian stiffness matrix for conservative systems and showed that the skew-symmetric part of a 6 × 6 stiffness matrix is related to the externally applied force/moment. Huang and Schimmels (1998) studied the spatial compliance behavior that can be achieved through the use of simple springs connected in parallel to a single rigid body. 3.1. Conservative Criteria of Stiffness Matrix For a stiffness matrix to be conservative, it must satisfy the following two conditions: (i) the force due to stiffness is integrable and conservative, and (ii) the work done by such force is conservative. For a linear stiffness matrix, K ∈ R3×3 , these two criteria can be used to deduce the following two conditions for conservative stiffness matrix (Kao and Ngo 1999). THEOREM 1. A robotic system with stiffness matrix is considered conservative if and only if 1. symmetry: The stiffness matrix K is symmetric: kij = kj i , and 2. exactness: The stiffness matrix K satisfies the exact dif∂k ik ferential condition: ∂xijk = ∂k ∂xj , where kij is the (i, j )-th element of the stiffness matrix and 1 ≤ i, j, k ≤ 3. An alternative proof of the above theorem using the tensor analysis is presented in Section A of the appendix. Criterion 1 is a well-known fact about stiffness matrix. Criterion 2 is a result of (curl K) = K× ∇ = 0; that is, if the external force f = K · dx is considered a point function, the exact ∂k ik differential condition ∂xijk = ∂k ∂xj should be satisfied, where ∂fi f is a function of independent variables and kij = ∂x . In j summary, Theorem 1 can also be stated as (1) (curl f)=0 and (2) (curl K)=0. This conservative property implies that the work done by a conservative force f depends only on the end states, not on the path. This is the basis of numerical integration in Section 4.2 to determine if a linear Cartesian stiffness is conservative. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 Chen and Kao / Conservative Congruence Transformation Both conservative criteria can also be applied to any stiffness matrix with generalized homogeneous coordinates. 3.2. Robot Stiffness and Conventional Congruence Mapping In robot grasping and manipulation, we can think of the grasped stiffness matrix, Kp , as a measure of interaction between the robot fingertip and object when external force is applied to the end-effector. By using the Taylor’s expan ∂f 1 ∂2f sion, we can write df = ∂x dx + 2 ∂x2 dx dx + · · · = ∂K Kp dx + 21 ∂xp dx dx + · · · , where df and dx are the force and displacement vectors in the Cartesian space relative to the robot’s base frame. If the higher-order small terms are neglected, the above equation is reduced to the familiar definition of stiffness matrix df = Kp dx or Kp = df . dx (2) Note eq. (2) represents the stiffness as an infinitesimal relationship.1 In addition, the manipulation Jacobian matrix, Jθ , relates the differential joint displacements to infinitesimal movement of the end-effector, i.e., dx = Jθ dθ, where dθ = [dθ1 , . . . , dθn ]T is the displacement of the joints. When we consider the forward kinematics of a serial robot manipulating an object under stiffness control, the grasped stiffness matrix associated with each manipulator can be derived from the structural compliance and the joint stiffness matrix Kθ (Cutkosky and Kao 1989). If the stiffness control is employed and the manipulator is always maintained at its unloaded configuration, the relationship between the joint and Cartesian stiffness matrices, defined in eq. (1), can be derived by the principle of virtual work (Salisbury 1980; Mason and Salisbury 1985). 3.3. Derivation of the Conservative Congruence Transformation From the principle of virtual work and the definition of the Jacobian, we obtain the following relationship between the joint torques and the force applied on the end-effector τ= JθT f. (3) Eq. (4) includes two separate terms. The first term is related to the changes in geometry; the second term is related to differential external force. Using the definition of stiffness in eq. (2), eq. (4) can also be written as ∂JθT T Jθ Kp dx = Kθ dθ − dθ f, (5) ∂θ dτ = (dJθT ) f + JθT (df). (4) 1. In some literature, the definition of df = −Kp dx is used, indicating that a restoring force is generated with the application of displacement. In this paper, we use the definition of df = Kp dx, where the increment of force df is seen as an addition to the external force f. Either definition will yield consistent overall results although certain terms in respective equations will differ by a sign. dτ df where the ∂JθT ∂θ and ∂JθT ∂θ dθ are third-order and second-order tensors (or matrices), respectively, and dx = Jθ dθ . Thus, we can simplify the first-order tensor, ∂JθT ∂θ dθ f, by the indicial notation (see Section B of the appendix for details) as follows: n ∂JT ∂JθT θ f dθi dθ f = ∂θ ∂θi i=1 n ∂JθT f dθi = ∂θi i=1 ∂JθT ∂JθT ∂JθT = ( f) ( f) · · · ( f) dθ. (6) ∂θ1 ∂θ2 ∂θn Eq. (5) can be rearranged to become ∂JθT ∂JθT ∂JθT T Jθ Kp Jθ = Kθ − ( f) ( f) · · · ( f) ∂θ1 ∂θ2 ∂θn (7) or JθT Kp Jθ = Kθ − Kg , (8) ∂JT where ( ∂θθi f) is an n × 1 column vector, with i = 1 . . . n, and n is the number of joints. The n × n matrix Kg , defining the changes in geometry via the differential Jacobian, is ∂JθT ∂JθT ∂JθT ∂JθT Kg = ( f) ( f) · · · ( f) = f , (9) ∂θ1 ∂θ2 ∂θn ∂θn n×n ∂JθT ∂θn f denotes an n × n matrix with the ith column ∂JθT element being ∂θi f and i = 1 · · · n (see Section B of the where It is an important force/torque relation that enables us to determine the resulting torques, τ , reflected at the joints as a result of the force applied at end-effector f, and vice versa. By the chain rule, we can differentiate eq. (3), 837 appendix). We call equation (8) the conservative congruence transformation (CCT). Therefore, the inverse of the CCT is Kp = Jθ−T (Kθ − Kg )Jθ−1 , (10) where the inverse is taken as the Penrose-Moore generalized inverse. Note that the conventional formulation in equation (1) did not consider the change of geometry resulting from the differential Jacobian and the applied load. Hence, it is Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 838 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / September 2000 true only at unloaded equilibrium configurations when f = 0. Since robot manipulators normally carry external load, equation (1) is not valid in general. From the CCT defined in eq. (8), we conclude the following corollaries. Proofs of Corollaries 3 and 4 are in Section C of the appendix. COROLLARY 1. When there is no externally applied force, i.e., f = 0, the conservative congruence transformation of eq. (8) becomes identical to the conventional formulation in eq. (1). Namely, eq. (1) is a special case of the conservative congruence transformation, and it is only valid at unloaded configuration of a robotic manipulator. COROLLARY 2. When the robot manipulator has a configuration-independent Jacobian, e.g., constant Jacobian matrix of a Cartesian gantry robot, the changes in Jacobian ∂JθT ∂θi are equal to zero, i.e., = 0. The conservative congruence transformation can be reduced to the conventional formulation. This is also a special case. COROLLARY 3. The conservative congruence transformation results in a symmetric mapping between the joint and Cartesian spaces. That is, the symmetry is preserved in the conservative congruence transformation and its inverse. COROLLARY 4. The positive definiteness of the stiffness matrices, Kp and Kθ − Kg , is preserved in both the conservative congruence transformation and the inverse congruence transformation. 3.4. Properties of Kg Matrix Next, we need to prove that the Kg matrix of the conservative congruence transformation defined in eq. (9) is also conservative based on the two criteria in Theorem 1. First of all, from eq. (9), we have Kg − KgT = T ∂JθT ∂JθT f − f . ∂θn ∂θn (11) Let us denote the (i, k)th element of Jθ as Jik . Thus, the (i, j )th element of (Kg − KgT ) is m ∂Jki k=1 ∂Jkj fk − f k ∂θj ∂θi m 2 ∂ xk ∂ 2 xk = fk = 0, − ∂θj ∂θi ∂θi ∂θj k=1 (12) where m is the number of degrees of freedom of f, m k=1 ∂Jki ∂θj fk represents the ith element of the j th column of Kg , ∂Jkj and m is the (i, j )th element of KgT . Therefore, k=1 fk ∂θi we conclude from eq. (12) that Kg is symmetric. Second, we can examine the exactness of the Kg matrix by ∂ ∂θl = m ∂Jki k=1 ∂θj m k=1 fk ∂ − ∂θj m ∂Jki k=1 ∂θl ∂ 3 xk ∂ 3 xk − ∂θl ∂θj ∂θi ∂θj ∂θl ∂θi fk (13) fk = 0. From eqs. (11) to (13), we conclude that the Kg matrix satisfies both symmetric and exact differential criteria; hence, Kg is conservative. Note that in the above proof, the generalized coordinates of the joint space, θ, satisfy the commutative property since they are independent parameters representing the displacements of each joint in the joint space. 3.5. Positive Definite Properties between Kθ and Kp in CCT We have proven that the positive definite properties are preserved between Kp and (Kθ − Kg ) in Section C of the appendix. While the effective stiffness matrix Kg is a function of the geometry and external force, a positive definite joint stiffness matrix, Kθ > 0, can be chosen or prescribed in stiffness control. The issue of concern here is the positive definite properties between the Kθ and Kp matrices. Since the Kg matrix depends on changes in geometry and external force, it can be positive definite or nonpositive definite. Based on eq. (10), we can rewrite eq. (38) as yT Kp y = zT Kθ z − zT Kg z, (14) where z = Jθ−1 y. Note that zT Kg z can assume positive or negative values, although we can prescribe positive definite joint control such that zT Kθ z > 0. Thus, it is obvious that the positive definiteness between Kθ and Kp is not necessarily preserved from eq. (14). We will discuss the following cases. (i) If (zT Kθ z − zT Kg z) > 0, then yT Kp y > 0. Thus, Kp is positive definite. Note that the magnitude and sign of zT Kθ z − zT Kg z may depend on the choice of z. For Kp to be positive definite, the above term has to be positive for any arbitrary nontrivial vector z. (ii) If (zT Kθ z − zT Kg z) ≥ 0, then Kp is positive semidefinite. (iii) If the external force and the geometry of the manipulator are arranged (for example, large external force applied in unfavorable direction) such that they render a Kg matrix that results in (zT Kθ z − zT Kg z) < 0, then Kp is nonpositive definite. That is, a positive definite joint stiffness control can result in a nonpositive definite Cartesian stiffness matrix due to external force and geometry that is manifested in the effective stiffness matrix Kg . This result suggests profound implication the external force and geometry could have on Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 Chen and Kao / Conservative Congruence Transformation 839 4.1. Analysis of Conservative Property of the CCT In this section, we will verify the conservative properties of the conservative congruence transformation using the joint-based control scheme, i.e., obtaining Kp of the Cartesian space with known Kg and Kθ in the joint space. We need to show, via the inverse congruence transformation of eq. (10), that the resulting Kp matrix from the known Kθ and Kg satisfies the two conservative criteria in Theorem 1. First of all, it is obvious that Kg is symmetric from eq. (15). Consequently, the resulting matrix Kp from eq. (10) is symmetric as long as Kθ is also symmetric. By substituting eq. (15) and a conservative jointstiffness3 into eq. (10), we k11 k12 obtain the components of Kp = , as follows: k12 k22 Fig. 1. A two-link planar robot manipulator. the stiffness control of a system. Similar effects are documented in the literature such as the “snap instability” due to large normal force in grasping with stiffness control (Cutkosky and Kao 1989). k11 = 4. Analysis and Simulation of a Two-link Planar Manipulator via the Conservative Congruence Transformation In this section, we will use a two-link planar manipulator to illustrate the conservative property of the Cartesian stiffness matrix, transformed from the joint stiffness matrix, via the CCT in eqs. (8) and (10) versus the conventional formulation in eq. (1). Numerical simulation will be provided to demonstrate the differences between them. In Figure 1, a two-link planar manipulator is illustrated with the end-effector at P (x, y) and the robot base frame at O(XY ). We define x = [x y]T and θ = [θ1 θ2 ]T for the planar robot. For the two-link manipulator, the stiffness matrix Kg defined in eq. (9) is ∂JθT −(L1 c1 + L2 c12 )fx − (L1 s1 + L2 s12 )fy Kg = f = (−L2 c12 fx − L2 s12 fy ) ∂θn (−L2 c12 fx − L2 s12 fy ) , (−L2 c12 fx − L2 s12 fy ) 2. The joint-based stiffness control is specified with a prescribed Kθ in the joint space. The Cartesian stiffness, Kp , is then obtained based on Kθ and Kg using the inverse CCT equation in (10). + 2 kθ11 − 2kθ12 + kθ22 c12 L21 s22 − kθ12 θ1 + kθ22 θ2 c12 c2 L22 s23 + 2 c kθ11 θ1 − kθ12 (θ1 − θ2 ) − kθ22 θ2 c12 2 L21 s23 + kθ11 θ1 − kθ12 (θ1 − θ2 ) − kθ22 θ2 c12 L1 L2 s23 2 kθ12 θ1 + kθ22 θ2 c12 L1 L2 s23 −kθ12 + kθ22 c12 s1 kθ22 c1 s1 + 2 2 2 L1 L2 s2 L2 s 2 − k12 = (15) where c1 = cos θ1 , s1 = sin θ1 , c12 = cos(θ1 + θ2 ), and s12 = sin(θ1 + θ2 ). Next, we will employ the conservative criteria to examine the stiffness matrices of the planar robot via the conservative congruence transformation. In addition, the numerical simulation of forces and work done under jointbased stiffness control2 scheme will be obtained and compared, based on both the CCT and conventional formulation. kθ22 c12 2(kθ12 − kθ22 ) c1 c12 − 2 2 L 1 L2 L2 s 2 s22 k21 = (16) − kθ12 θ1 + kθ22 θ2 c1 c2 s1 L22 s23 + kθ11 θ1 − kθ12 (θ1 − θ2 ) − kθ22 θ2 c1 s1 L1 L2 s23 − kθ12 − kθ22 c1 s12 kθ11 − kθ12 + kθ22 c12 s12 + 2 L1 L 2 s2 L21 s22 + kθ11 θ1 − kθ12 (θ1 − θ2 ) + kθ22 θ2 c2 c12 s12 L21 s23 − kθ12 θ1 + kθ22 θ2 c12 s12 L1 L2 s23 k12 3. We use a symmetric and constant Kθ = convenience of illustration. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 (17) (18) kθ11 kθ21 kθ12 kθ22 matrix for 840 k22 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / September 2000 = kθ22 s12 kθ12 θ1 + kθ22 θ2 c2 s12 − L22 s22 L22 s23 In this simulation, the initial joint torque is given as τ0 = [10 20]T N · m, resulting in the initial bias force of f0 = [73.1 96.7]T N at the initial configuration of θ1 = 1.21rad and θ2 = −1.09rad. The force and work done in the Cartesian space are computed by employing the two-dimensional trapezoidal rule + kθ11 θ1 − kθ12 (θ1 − θ2 ) − kθ22 θ2 s12 L1 L2 s23 − 2 2(kθ12 − kθ22 ) s1 s12 kθ11 − 2kθ12 + kθ22 s12 + L1 L2 s22 L21 s22 + 2 kθ11 θ1 − kθ12 (θ1 − θ2 ) + kθ22 θ2 c2 s12 L21 s23 − 2 kθ12 θ1 + kθ22 θ2 s12 . L1 L1 s23 fi = fi−1 + Kp dx (22) 1 Wi = Wi−1 + dW = Wi−1 + fiT dx + dxT Kp dx. 2 (23) (19) Thus, the exactness (i.e., criterion 2 in Theorem 1) of the Cartesian stiffness matrix can be examined by substituting eqs. (16) through (19) into the following equations: ∂k11 ∂k12 ∂k11 ∂θ2 ∂k11 ∂θ1 − = + ∂y ∂x ∂θ1 ∂y ∂θ2 ∂y ∂k12 ∂θ1 ∂k12 ∂θ2 − + =0 (20) ∂θ1 ∂x ∂θ2 ∂x ∂k21 ∂k22 ∂k21 ∂θ2 ∂k21 ∂θ1 − = + ∂y ∂x ∂θ1 ∂y ∂θ2 ∂y ∂k22 ∂θ1 ∂k22 ∂θ2 − + = 0, (21) ∂θ1 ∂x ∂θ2 ∂x ∂θi −1 i where the terms ∂θ ∂x and ∂y are the entries of the Jθ matrix. Eqs. (20) and (21) can be verified by symbolic software such as Mathematica or Maple V. From the above analysis, we conclude that the Cartesian stiffness matrix transformed from a constant and symmetric joint stiffness matrix via the CCT is conservative because it satisfies both criteria of symmetry and exactness. 4.2. Numerical Simulation of Congruence Transformation In the following sections, we will present and compare the results of numerical simulation of both congruence transformations. Since Kg is shown to be conservative in Section 3.4, we consider in the following sections three cases: (i) conservative Kθ , (ii) nonconservative Kθ , and (iii) special case. 4.2.1. Case 1: Conservative Kθ We will use the example of the two-link manipulator in Figure 1 with link lengths, L1 = 0.29m and L2 = 0.23m, under joint-based stiffness control with a prescribed conservative joint stiffness matrix 10 0 Kθ = N · m/rad. 0 20 An alternative derivation of dW in eq. (23) is in Section D of the appendix. The last quadratic term in eq. (23) usually has a smaller contribution than the second term. The path of integration is a circle with a radius of 0.08m centered at (x, y) = (0.25, 0.30)m, starting from the initial end-effector position at x0 = 0.33m and y0 = 0.3m, tracing a counterclockwise sense. If the system is conservative, we expect that the net work done over the closed circular path is zero and the net change in force is also zero. Similarly, the work done in the joint space is calculated by 1 Wi = Wi−1 + dW = Wi−1 + τiT dθ + dθ T Kθ dθ. 2 (24) Note that the work done in the Cartesian and joint spaces is only integrated from the force/torque and Kp /Kθ , respectively. The difference between the two congruence transformations is that Kp in the CCT is obtained from both Kθ and Kg . The results of numerical simulation are shown in Figure 2(a) and 2(b) for the conventional formulation and CCT, respectively. A few observations are in order: • In Figure 2(a), the conventional formulation yields net changes in forces after integrating over the closed path, i.e., %fx = 0 and %fy = 0. This shows that the conventional formulation is not a conservative mapping. • In addition, the net work in the Cartesian space for the conventional formulation in Figure 2(a) is not zero, whereas the net work done in the joint space is zero. The net work calculated in the joint space is zero because this is a joint-based control scheme (Kao and Ngo 1999), and the work is conservative as long as both the Kθ and Kg matrices are conservative. However, the Kp matrix is obtained through the nonconservative mapping of eq. (1), thus resulting in discrepancy in the work done. It is obvious that the two trajectories for work done in Figure 2(a) should have been the same. This further shows that the conventional formulation does not yield correct results when external force is applied. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 Chen and Kao / Conservative Congruence Transformation 841 Fig. 2. Simulation of joint-based control system with a conservative Kθ : The parameters are given in Section 4.2.1. The plots show the Cartesian force f = [fx fy ]T and the work done over the closed circular path in counterclockwise sense for (a) the conventional formulation and (b) the conservative congruence transformation (CCT). • The profiles of forces in Figure 2(a) are erroneous due to the employment of the conventional formulation in calculating the Cartesian forces. nonconservative joint stiffness matrix via the conventional formulation and CCT, respectively. We observe the following: • When applying the conventional formulation, the discrepancy in work will become more pronounced when the external force is larger. The magnitude of this discrepancy also depends on the stiffness matrix and the differential Jacobian. • Since this is a nonconservative system with a nonconservative Kθ , we expect that the net work done is nonzero. Figure 3 clearly shows the net works are not zero. • In Figure 2(b), the CCT yields identical work profiles with zero net work in both joint and Cartesian spaces and results in zero changes in forces over the closed path, i.e., dW = 0 and df = 0. There is no discrepancy in work done, showing that the CCT yields the exact results and is a conservative mapping between the joint and Cartesian spaces. 4.2.2. Case 2: Nonconservative Kθ Contrary to the previous section, we will use an asymmetric (nonconservative) joint stiffness matrix to perform the simulation. First of all, we know both congruence mappings will result in an asymmetric Cartesian stiffness matrix, even though the Kg matrix is still a conservative matrix (Section 3.4). Thus, we obtain a nonconservative Cartesian stiffness matrix transformed from a nonconservative joint stiffness via both congruence transformations. The nonconservative joint stiffness matrix used in the simulation is 10 5 Kθ = N ·m/rad. 0 20 The setup of simulation is the same as those in Section 4.2.1. Figures 3(a) and (b) present the simulation results of the above • Similar to Figure 2(a), Figure 3(a) gives incorrect results for the force components, fx and fy . • Again, the work profiles of Figure 3(b) are identical, indicating that the CCT is a correct mapping between the joint and Cartesian spaces. On the other hand, Figure 3(a) shows the discrepancy in works done. • The discrepancy of work profiles for Figure 3(a) becomes more pronounced as the external force becomes larger. • It is interesting to note that the net change in force in Figure 3(b) is zero even though the system is nonconservative. This is because the exactness criterion in Theorem 1 is met; hence, the CCT under this circumstance will result in %fx = %fy = 0. On the contrary, Figure 3(a) yields %fx = 0 and %fy = 0, which violates the expectation because both Kθ and Kg satisfy the exactness criterion. 4.2.3. Case 3: Special Configuration-Dependent Kθ In addition to the two cases in Sections 4.2.1 and 4.2.2, we consider the special Kθ in Kao and Ngo (1999). In the analysis of Kao and Ngo and the simulation of Chen and Kao (1998), Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 842 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / September 2000 Fig. 3. Simulation of joint-based control system with an asymmetric and nonconservative Kθ : The parameters are given in Section 4.2.2. The plots show the Cartesian force f = [fx fy ]T and the work done over the closed circular path in counterclockwise sense for (a) the conventional formulation and (b) the conservative congruence transformation (CCT). they presented a special form of configuration-dependent Kθ matrix for a two-link manipulator that renders conservative Kp matrix via the conventional formulation. When L1 = L2 , the special joint stiffness matrix is chosen as 2k k Kθ = , cos θ2 k 1+cos θ2 k where k is a constant and is taken as 7.5 in the following simulation. Figures 4(a) and 4(b) show the results of simulation, with L1 = L2 = 0.29m via the conventional formulation and CCT, respectively. Other parameters and conditions are held the same as those in Sections 4.2.1 and 4.2.2. The results show that while both Figures 4(a) and 4(b) yield the same end states (i.e., %fx = %fy = %W = 0), there is a difference between the profiles. In particular, the discrepancy in work profiles exists in Figure 4(a). Nevertheless, the CCT still yields identical work profiles and correct results, as shown in Figure 3. 4.3. Summary We can summarize the results from the proceeding analysis and numerical simulation. 1. The conventional formulation always produces discrepancies between the profiles of work done in the Cartesian and joint spaces. This is because the conventional mapping is not a conservative mapping (Kao and Ngo 1999) and does not yield correct results unless the grasp is always maintained at its unloaded configuration. 2. The discrepancies of force and work between the joint and Cartesian spaces of the conventional formulation become more pronounced as the influence of the external force becomes larger. 3. In the CCT, the work done in both joint and Cartesian spaces is always identical and consistent. Thus, the CCT is a correct and complete transformation of stiffness matrices between the joint and Cartesian spaces. 4. In addition to the consistency of the conservative property in net work, the relation between torque and force is also verified in the CCT. The simulation results are shown in Figure 5, where the parameters of simulation for case (1), with a conservative Kθ in Section 4.2.1, are used. In the figure, we compute the torque, τ1 and τ2 , based on the numerical simulation in the joint space, as well as the torque calculated from the Cartesian force using τ = JθT f. We expect that the two sets of torques should be identical if the transformation equation between the Kθ and Kp is correct. It is clear from the figure that the torques computed from the CCT are consistent and identical. On the other hand, the conventional formulation always yields erroneous results in both Cartesian forces and torques calculated from forces in the joint-based stiffness control system. 5. It is essential that the conservative congruence transformation is employed whenever external force is expected. 6. The Kg matrix captures the effect due to the changes in geometry under the presence of external force. It also accounts for the discrepancy of work profiles in the simulation of conventional formulation. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 Chen and Kao / Conservative Congruence Transformation 843 Fig. 4. Simulation of joint-based control system with the special Kθ (Kao and Ngo 1999): The parameters are given in Section 4.2.3. The plots show the Cartesian force f = [fx fy ]T and the work done over the closed circular path in counterclockwise sense for (a) the conventional formulation and (b) the conservative congruence transformation (CCT). Fig. 5. Simulation results of torque integrated directly in the joint-based stiffness control system, and the torques calculated from the Cartesian force using τ = JθT f, where Jθ is updated for each configuration in the simulation. The parameters used are those of Case 1 in Section 4.2.1. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 844 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / September 2000 7. The graphical illustration of the CCT is offered in Figure 6, where Kθ in the joint space is combined with the effective stiffness matrix, Kg , for the conservative mapping. 8. The elbow-out configuration of the manipulator was used in the simulation, as shown in Figure 1. The CCT can be equally applied to the elbow-in configuration, with different numerical results. And yet, the conclusions of the conservative properties of force and work remain the same. 5. Stiffness Control of Robot Grasping Manipulation via the Conservative Congruence Transformation From the preceding analysis, we find that the conventional formulation in eq. (1) is a special case of the conservative congruence transformation when the system is always maintained at its unloaded equilibrium configuration. Equations can be formulated for displacements and force/torque by tracing the matrix terms in square boxes. If the direction of an arrow is followed, the matrix term in the box will be used, whereas if the arrow is traced in the reversed direction, the inverse of the matrix term should be employed. Figure 7 illustrates the relationship of the stiffness matrices between the joint and Cartesian spaces based on the conservative congruence transformation. From Figure 7, we find the relation between displacements of the joints, dθ, and the changes of externally applied forces, df, as df = Jθ−T Kθ − Kg dθ. (25) In addition, we can write the relationship between the changes in the joint torques, dτ , and the changes in the Cartesian forces, df, as df = Jθ−T Kθ − Kg Kθ−1 dτ (26) = Jθ−T dτ − Jθ−T Kg Kθ−1 dτ. The first term of eq. (26), df = Jθ−T dτ , relates the infinitesimal changes of df and dτ when there is no initial bias force/torque. Nevertheless, when relating the changes in torques/forces in the joint/Cartesian spaces, the relationship of eq. (26) should be used to include the effect of load and changes in geometry. Other relationships of stiffness control and mapping between the joint and Cartesian spaces can be traced using Figure 7, along with eqs. (25) and (26). between the joint and Cartesian spaces and should be employed to replace the conventional formulation that is only valid at the unloaded equilibrium configuration. The conservative congruence transformation accounts for the effect of change of robot grasping geometry under the presence of external force. It is shown that the Cartesian stiffness must be obtained from two parts: one from the active joint stiffness control Kθ and the other from the effective stiffness matrix resulting from the changes in geometry as represented by the Kg matrix. The conservative congruence transformation also preserves the symmetric, positive-definite, and conservative properties of stiffness matrices when mapping between the Cartesian and joint spaces. Appendix Alternative Derivation for the Conservative Criteria of the Stiffness Matrix Here, we employ the indicial notation (Malvern 1969) to derive the conservative criteria for the stiffness matrix (Ngo 1998). We consider the external force f as a vector field, and the curl of the vector field f is denoted as (curl f). If the vector field f is the gradient of a scalar function, then (curl f)=0. Conversely, it can be proved that if (curl f)=0 throughout a simply connected region, then f is the gradient of a potential function φ defined by T φ = f dr = f1 dx1 + f2 dx2 + f3 dx3 , (27) C C which is a line integral along any curve C from an arbitrary initial point in the region to a different destination point of the region. The necessary and sufficient conditions for φ to be a point function is (curl f)=0, which makes the line integral of eq. (27) independent of path on a simply connected region in which f and (curl f) are continuous. In indicial notation, the curl operator from the right on a vector given in rectangular Cartesian space yields f × ∇ = (fr ir ) × ( ∂ p ip ) = fr ∂ p (ir × ip ) where erpq is the alternating (or permutation, or Levi-Civita) ∂f ∂fi symbol. Thus, (curl f) = 0 implies kij = ∂x = ∂xji = kj i . j Next, we need to prove that the force, f = K · dx, is also a point function. The curl operator from the right for a second-order tensor yields K × ∇ = (krs ir is ) × ( ∂ p ip ) = krs ∂ p ir (is × ip ) 6. Conclusions The conservative congruence transformation (CCT) presented in this paper is the correct and complete stiffness mapping (28) = erpq fr,p iq , = espq krs,p ir iq , with typical components: Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 (29) Chen and Kao / Conservative Congruence Transformation Fig. 6. Graphical illustration of the conservative congruence transformation and its inverse. Fig. 7. Relationships of stiffness control in robotic grasping and manipulation. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 845 846 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / September 2000 (K× ∇ )11 (K× ∇ )12 (K× ∇ )13 (K× ∇ )21 (K× ∇ )22 (K× ∇ )23 (K× ∇ )31 (K× ∇ )32 (K× ∇ )33 = ∂k12 ∂k13 − ; ∂x3 ∂x2 = ∂k13 ∂k11 − ; ∂x1 ∂x3 = ∂k11 ∂k12 − ∂x2 ∂x1 = ∂k22 ∂k23 − ; ∂x3 ∂x2 = ∂k23 ∂k21 − ; ∂x1 ∂x3 = ∂k21 ∂k22 − ∂x2 ∂x1 = ∂k32 ∂k33 − ; ∂x3 ∂x2 = ∂k33 ∂k31 − ; ∂x1 ∂x3 = ∂k31 ∂k32 − . ∂x2 ∂x1 where dθk and fj are first-order tensors. On the right-hand side of the arrows, we draw the equivalence of the tensor analysis to the terms used in Section 3.3. Thus, eq. (35) can be rewritten as n m ci = aij k dθk fj j =1 (30) k=1 = (31) (32) (33) We can define a second-order tensor as ∂JθT dθ, ∂θ where aij k represents the components of ∂JθT ∂θ (34) and bij rep ∂JT resents the components of second-order tensor in ∂θθ dθ . The symbol “⇐⇒” is used to denote equivalence. Similarly, for a first-order tensor, we can write m ∂JθT ci = bij fj ⇐⇒ dθ f, (35) ∂θ j =1 aij 1 fj dθ1 + · · · + m aij n fj dθn j =1 ⇐⇒ ∂JθT f dθn , ∂θn (36) Proofs of Corollaries 3 and 4 Derivations for the CCT k=1 where n = number of joints, m = number of degrees of freedom in the Cartesian space, and aij 1 . . . aij n are the entries of the third-order tensor, aij k . for 1 ≤ i, j, k ≤ 3. aij k dθk ⇐⇒ m The curl operator from the left will give rise to the same results as above (see Ngo 1998). Summary: (curl f)=0 and (curl K)=0 are the criteria for conservative stiffness. n j =1 j =1 bij = n m = aij k fj dθk It is obvious that for all of the components of (K× ∇ ) to be equal to zero, we must have ∂kij ∂kik = ∂xk ∂xj k=1 Proof of Corollary 3. For joint-based stiffness control, Kg is symmetric based on eqs. (11) and (12), and Kθ is symmetric in a conservative stiffness control system. The resulting stiffness matrix in the Cartesian space can be found using the inverse CCT in eq. (10). Hence, T T KpT = Jθ−1 KθT − KgT Jθ−T = Jθ−T (Kθ − Kg )Jθ−1 (37) = Kp . That is, Kp is also symmetric. Similarly, we can prove that the symmetry is preserved for the Cartesian-based stiffness control by using the CCT in eq. (8). Proof of Corollary 4. First, let us regard (Kθ − Kg ) as one entity to relate with Kp . Next, we multiply both sides of eq. (10) by an arbitrary vector y to render yT Kp y = yT Jθ−T (Kθ − Kg )Jθ−1 y = zT (Kθ − Kg )z, (38) where z = Jθ−1 y. The vector y is arbitrary and so is z. From eq. (38), we can find that if Kp is positive definite (i.e., yT Kp y > 0), then zT (Kθ − Kg )z > 0, which means that (Kθ − Kg ) is also positive definite. On the other hand, if (Kθ − Kg ) is positive definite, then Kp will also be positive definite. This proof can be extended to positive semidefinite matrices. Therefore, it is clear that the positive definite properties are preserved between Kp and (Kθ − Kg ). Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016 Chen and Kao / Conservative Congruence Transformation Equation of Work Integration Here, we present a different derivation for eq. (23) using the Taylor’s expansion. For stiffness controlled system, the potential energy is the work. Therefore, the work of this stiffness system is W = U . We can write dU = ∂U 1 ∂ 2U dx + dxT dx + · · · . ∂x 2 ∂x2 We note that f T = ∂U ∂x and K = order small terms, we have ∂2U . ∂x2 (39) Neglecting the higher- 1 dW = f T dx + dxT K dx, 2 (40) which renders the same equation as (23), which was presented using the more well-known numerical algorithm (trapezoidal rule). Acknowledgments The research was supported by the NSF/ARPA Grant IRI9309823 and IIS9906890. References Ball, R. S. 1900. A Treaties on the Theory of Screws. Cambridge, UK: Cambridge University Press. Chen, S.-F., and Kao, I. 1998. Simulation of conservative properties of stiffness matrices in congruence transformation. Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, Victoria, BC, Canada, pp. 311–316. Chen, S.-F., and Kao, I. 1999. The conservative congruence transformation of stiffness control in robotic grasping and manipulation. The 9th International Symposium on Robotics Research, Snowbird, UT, pp. 7–14. Ciblak, N., and Lipkin, H. 1994. Asymmetric Cartesian stiffness for the modeling of compliant robotic systems. Robotics: Kinematics, Dynamics and Control, ASME DE Vol. 72, pp. 197–204. Cutkosky, M. R., and Kao, I. 1989. Computing and controlling the compliance of a robotic hand. IEEE Transaction of Robotics and Automation 5(2):151–165. Dimentberg, F. M. 1965. The screw calculus and its applications in mechanics. Technical Report FTD-HT-23-163267, U.S. Department of Commerce Translation. 847 Griffis, M., and Duffy, J. 1993. Global stiffness modeling of a class of simple compliant couplings. Mechanism and Machine Theory 28(2):207–224. Hogan, N. 1990. Mechanical impedance of single- and multiarticular systems. In Multiple Muscle Systems: Biomechanics and Movement Organization, ed. J. M. Winters and S. L.-Y. Woo, 149–164. New York: Springer-Verlag. Howard, S., Žefran, M., and Kumar, V. 1998. On the 6 × 6 stiffness matrix for three dimensional motions. J. of Mechanism and Machine Theory 33(4):389–408. Huang, S., and Schimmels, J. M. 1998. The bounds and realization of spatial stiffnesses achieved with simple springs connected in parallel. IEEE Transaction on Robotics and Automation 14(3):466–475. Kao, I., Cutkosky, M. R., and Johansson, R. S. 1997. Robotic stiffness control and calibration as applied to human grasping tasks. IEEE Transaction on Robotics and Automation 13(4):557–566. Kao, I., and Ngo, C. 1999. Properties of grasp stiffness matrix and conservative control strategy. The International Journal of Robotics Research 18(2):159–167. Lončarić, J. 1985. Geometrical Analysis of Compliant Mechanisms in Robotics. Ph.D. dissertation, Harvard University, Cambridge, MA. Malvern, L. E. 1969. Introduction to the Mechanics of a Continuous Medium. Englewood Cliffs, NJ: Prentice Hall. Mason, M. T., and Salisbury, J. K. 1985. Robot Hands and the Mechanics of Manipulation. Cambridge, MA: MIT Press. Mussa-Ivaldi, F. A., Hogan, N., and Bizzi, E. 1985. Neural, mechanical, and geometric factors subserving arm posture in humans. Journal of Neuroscience 5(10):2732–2743. Ngo, C. 1998. Conservative Properties of Grasp Stiffness Matrix. Master’s thesis, Department of Mechanical Engineering, State University of New York at Stony Brook. Pigoski, T., Griffis, M., and Duffy, J. 1992. Stiffness mapping employing different frames of reference. Proceedings of the 22nd Biennial ASME Mechanisms Conference, ASME, Phoenix, AZ, pp. 445–452. Salisbury, J. K. 1980. Active stiffness control of a manipulator in Cartesian coordinates. Proceedings of the 19th IEEE Conference on Decision and Control, Albuquerque, NM, pp. 87–97. Žefran, M., and Kumar, V. 1997. Affine connections for the Cartesian stiffness matrix. Proceedings of the IEEE International Conference on Robotics and Automation, Albuquerque, NM, pp. 1376–1381. Downloaded from ijr.sagepub.com at PENNSYLVANIA STATE UNIV on September 19, 2016
© Copyright 2026 Paperzz