2011 IEEE/RSJ International Conference on Intelligent Robots and Systems September 25-30, 2011.

San Francisco, CA, USA

Impedance control of a non-linearly coupled tendon driven thumb

Maxime Chalon, Werner Friedl, Jens Reinecke, Thomas Wimboeck, Alin Albu-Schaeffer
Institute of Robotics and Mechatronics, German Aerospace Center (DLR), Wessling, Germany E-mails: {Maxime.Chalon, Werner.Friedl, Jens.Reinecke}@dlr.de

Abstract A large workspace and proper force capabilities of a robotic thumb can be obtained using a tensegrity structure for the actuation, similar to the human thumb base muscles. Using nonlinear stiffness elements and an antagonistic architecture, the joint stiffness can be adjusted by variation of the tendon pre-tension. However, the highly nonlinear actuation creates new control challenges and in particular the nonlinear tendon kinematics must be accounted for. Despite the challenges, the nonlinear structure is required to achieve the desired torques. In this paper, the dynamic equations of a tendon driven thumb are established. An efcient formulation is proposed to generate the pretension forces in order to preserve the torques and approximate the stiffness matrix. A cascaded structure is used for the controller. The equations for the inner tendon force control loop and the outer impedance control loop are presented. Because of the absence of link side position sensors, an iterative estimation algorithm is proposed and implemented in real-time. It is shown that, using the mechanical joint exibility, the controller impedance gain can be adjusted to improve the steady-state effective impedance. The search algorithm robustness is evaluated through a set of simulations. Finally, experimental results and equivalent simulations demonstrate the effectiveness of our controller.

Fig. 1. Hand Arm System with 19 DOF driven by 38 motors. The thumb has 4 DOF driven by 8 tendons. The four base tendons have a tensegrity structure.

I NTRODUCTION Numerous robot hand designs are based on tendon driven systems. The remote actuation provides low link inertia, small form factor for the link side and high dynamics. Moreover, in conjuction with variable stiffness elements in an antagonistic conguration, it is possible to control position and stiffness of the joints independantly. In [1], Grebenstein et al. highlight that the intrinsic compliance seen as a drawback in the past, can be a feature today. For small robots, such as hands, it is especially interesting in order to protect the robot itself. Indeed, since a control law can not react in the very rst instant of the impact, the only protection of the robot is its intrinsic compliance. In [2], [3], the special role of the thumb was studied in order to build an anthropomorphic hand arm system that would, from the design on, provide a functional thumb. It appeared that to generate sufcient thumb forces a special tensegrity structure of the tendons was required. However, the work did not address the control issues associated with induced non-linearities. In the developed system, the tendons are pulled by remote actuators (brushless motors). Flexible tendons are used in an antagonistic conguration such that one part of the tendon forces generates joint torques and the other part, known as pretension forces, adjusts the intrinsic stiffness of the joints. Examples of mechanisms used to obtain variable stiffness behavior can be found in [4][8].
978-1-61284-456-5/11/$26.00 2011 IEEE

Robot hands with a 2N actuation scheme joint have been proposed (2 tendons attached to each). To reduce the tendon count, hands with N joints driven by N + 1 were designed. Several controllers were presented that are capable of maintaining the tendon tension while moving [9][11]. A controller for a variable stiffness mechanism using two motors and two springs in an antagonistic conguration is presented in [12]. Feedback linearization of uncoupled joints that are each driven antagonistically was treated in [13]. Adaptive motor level PD controller was presented in [8] and a neural network based controller for exible tendon was proposed in [14] . More recently the implementation and control of the Robonaut tendon driven ngers were presented in [7]. We extensively studied passitivity based impedance controllers in [15], [16]. An impedance control scheme for variable stiffness mechanisms with nonlinear joint coupling was proposed in [17] but provided only simulation results. In the eld of grasping the use of soft ngertips is proposed to generate variable stiffness for a grasp by modication of the internal grasp force [18]. However, the analysis was limited to the ngertips and was static. [19], [20] proposed designs of variable compliance actuators for robotic subsystems and addressed the problem of choosing the appropriate joint compliance in order to attain a desired Cartesian compliance. In order to achieve a large range of motion and/or to minimize frictional effects the use nonlinear tendon routing was proposed. In this case, not only the compliance, but at the same time the tendon routing and hence the joint coupling becomes nonlinear. In this work, the very specic conguration of the tendon of the thumb base creates new challenges. Because of the

4215

geometric changes the coupling matrix (with relates the tendon velocities to the joint velocities) is not position independant. The main contribution of this paper is the implementation of an impedance controller for tendon driven robot thumb of the DLR Hand Arm System, that has to cope with nonlinear (i.e. conguration dependant) coupling matrix and the adjustment of the mechanical joint stiffness. Since the joint position is not measured, a novel method is proposed to estimate it based on a projected gradient descent approach. These concepts are veried by the experimental evaluation. In the rst section the general modeling of the thumb is proposed. The tendon kinematics are derived from the CAD model. In the second section, an impedance controller is derived. It is shown that the mechanical stiffness adjustement requires to modify the tendon pretensions. Section three describes the method used to estimate the link side position. The fourth section formulates an optimisation problem that selects the proper tendon forces in order to obtain the desired mechanical stiffness as well as exactly producing the impedance controller torques. Section ve presents a method adjust the controller stiffness, depending on the mechanical stiffness, such that the effective impedance is achieved. Section six presents the simulation and experimental results. The robustness of the gradient search algoritm is studied. The close loop behavior is demonstrated through a set of simple examples. I. T HUMB M ODELING AND C ONTROL In this section, the fundamental dynamic equations for the thumb are derived. Using the CAD data, a model based dynamical system is established. The thumb Fig. 2 has 4 degrees of freedom (DOF) that are driven by 8 tendons. The base joint is driven by four tendons are placed in a pyramidal conguration to increase the moment arm. The PIP and DIP joints are driven by four more tendons. The latter tendons are nearly going throught the rotation center of the MCP1 , and therefore have a neglectable inuence on the base torques. q4 q3

TABLE I D ESCRIPTION OF VARIABLES Symbol nN mN h Rm q Rn f Rn f Rm Kt Rmm Kq Rnn kti (fti ) R P Rnm M Rnn B Rmm C (q, q) g (q ) Rn f ric,q Rn m Rm Designation number of mechanical degree of freedom (joints) number of tendons positions of the tendon w.r.t. a xed reference positions of the joint joint torques generated by the tendon forces forces applied on the tendon stiffness matrix of the tendons stiffness matrix of the joints stiffness of the tendon depending on the tendon force coupling matrix link side mass matrix diagonal motor side mass matrix link side centrifugual and coriolis terms link side gravity vector frictional torque link motor torque

of generic variable stiffness mechanism, with exible joints are [17]:

M (q ) q + C (q , q ) + g (q ) + f ric,q + f ric, + Ef (q , ) B = P (q )f (q , ) + ext (1) = m

E is a m m diagonal matrix of the motor pulley radii, i.e. i [1..m], E (i, i) = ri . The coupling matrix P (q ) Rm to the Rnm , which relates the m tendon velocities h n n joint velocities q R . It is obtained as the derivative of the tendon position h with respect to the joint position q . As usual, is the vector of the motor positions. P (q ) = h(q ) q
T

(2)

Using the principal of virtual work it yields: q = P (q )f (q , ) (3)

The tendon forces depend on the tendon elongation. Therefore, the stiffness of a tendon is locally obtained as: kt (q 0 , 0 ) = f (q , ) |q=q0 ,=0 h(q , ) (4)

Since, in the considered design the tendon are independant, the tendon stiffness form a diagonal matrix of the tendon stiffness K t Rmm (5) h4 q1,2 h3 It is important to note that the equations are holding only as long as the tendon forces are in their prescribed working range f [fmin ..fmax ]. In the rest of the paper this assumption is considered true. II. T HUMB I MPEDANCE C ONTROLLER h5,6,7,8 In this section, an impedance controller for a tendon actuated mechanism is presented. The decoupling between link side and motor side is solved by using a cascaded structure. An inner tendon force controller tracks the desired tendon forces and guarantees that the tendons are never overstretched or loose. The tendon force controller also

h2 h1
Fig. 2.

Four joints and their associated 8 tendons

Table I denes the used symbols. The dynamic equations

1 Proximal Inter Phalangeal, Distal Inter Phalangeal and Metacarpal Proximal

4216

naturally compensate the tendon creeping (for more details on the tendon behavior see [21]). An outer joint impedance controller creates the desired link behavior. The estimation of the link positions, required for the impedance controller, is performed via a gradient search algorithm. A. Tendon force controller From eq. (2) it appears that only part of the tendon forces are generating a torque. Therefore, it is possible to adjust the joint stiffness and the joint torque by setting the tendon pretension forces and the tendon forces difference. Using eq. (3) the tendon forces can be decomposed in two parts: f t = P + (q ) q + f int (6)

B. Impedance Controller Based on the cascaded controller hypothesis, one can design the link side controller considering the joint torque created by the tendon forces, des , as a control input. Indeed, replacing eq. (8) in eq. (1) yields: M (q ) q + C (q , q ) + g (q ) + f ric,q = P (P EE
+ 1

des + (I P P + f intuser ) + ext (11) = des + ext

The force tracking errors terms are neglected for clarity. The targeted link dynamics are those of an impedance behavior. Namely: + K de + K pe = 0 M imp e (12)

Where P + (q ) is the generalized pseudo inverse, link is the desired link torque and f int is a force vector in the null space of P (q ) (further denoted N (P )). The pseudo inversion can be weighted such that the solution minimize a specic cost function [22]. The non weighted Moore-Penrose pseudo inverse minimizes the internal forces and distributes the forces equally (in the least mean squares sense). It is not, in general, possible to specify f int directly since it must lie into the nullspace of P (q ) (to avoid creating undesired torques) which is time varying. Therefore, user dened pretension forces f intuser can simply be projected into the nullspace of P (q ), to avoid inuence on the joint torque, via: f int = (I P (q )P + (q ))f intuser (7)

Where e, the position error, is dened as e = q q des . The link side controller input is consequently designed as: f ric,q des = imp + With imp being the desired impedance law: , q des + C (q imp = M q ) + g (q ) + K p (q des q ) (14) ) +K d (q des q Replacing eq. (13) and eq. (14) in eq. (1) yields: + K de M imp e + K p e = K tau 1 (K d q + K pq ) (15) (13)

Similar to the work in [23], the reference tendon torques are designed as: t,ref = E (P (q )+ des + (I P (q )P + (q ))f intuser ) (8) The structure of eq. (8) guarantees that the user dened tendon pretension has no inuence on the produced joint torque. Moreover, the joint torque is achieved exactly. However, this approch does not guarantee that the resulting tendon forces are satisfying the pulling constraints. Indeed, the user pretension forces, once projected may not create a sufcient force offset. This fundamental issue is addressed in the section IV. The tendon force controller is designed as a proportional controller with feedforward terms and friction compensation: m = f ric, + t,ref + K ( t,ref ) (9)

Where ( .) denotes the error between an estimated quantities and the real one (for clarity C (., .), f ric,q and g (.) are omitted). It should be noted that eq. (15) corresponds to the desired impedance behavior as long as the error between the estimated joint position and the real joint position is small. The error mainly depends on the mechanical joint stiffness since the link position is estimated only based on the motor positions (in order to preserve passivity). III. L INK POSITION ESTIMATION The ngers of the Hand Arm System have no link side sensors, confering them an astonishing robustness [24], therefore it is necessary to estimate the position of the joints. The link position can be estimated using a least square t based on the redundant tendon measurements. However, only the motor position should be used and not the force sensor deection to ensure passivity. In the case of a linear coupling, the = Pq relation h between joint velocity and tendon velocities (motor velocities) can simply be integrated to obtain the link position, for example, with the following pseudo inversion. q = P + h + q0
n

Where is the estimated tendon torques based on the tendon forces. The tendon torques t are obtained from the tendon forces with t = rpulley f t . Replacing eq. (9) in eq. (1) leads to: + ( t t,ref ) = K ( t B ) (10)

(16)

In pratice, the motor inertia term can be neglected and the torque feedback gain can be large. Therefore, the tendon force controller ensures that t,ref can be tracked.

Where q 0 R is some initial position reference. However, in the case of the thumb, the coupling matrix is non linear and cannot be integrated (the numerical time integration tends to drift). To estimate the joint position q , a non linear optimisation problem can be formulated: min( h(q ) hmeasured )
q

(17)

4217

The robustness of the algorithm, with respect to the tendon displacement error, is important. Fig. VI-A shows the algoritm performance on the thumb base range of motion depending on the number of iterations and the (simulated) noise on the tendon displacement. A classical gradient search algorithm is used. The algoritm is also written to t in the realtime process, and therefore, has a constant number of steps. In the online implementation, the previous result, if valid, is used as starting point. If q is out of range, the previous valid value is kept and the next starting point is set to 0. If the next step fails, the realtime program stops. IV. T ENDON PRETENSION CONTRAINTS AND JOINT
MECHANICAL STIFFNESS

arm system the stiffness model is best approximated by a polynom (Fig. 3), therefore, it is not possible to explicitly nd the inverse function 1 that maps the desired joint torques and stiffness to the tendon forces. However, if a suitable tendon stiffness model is selected, the inversion can be reduced to a simple matrix inversion and allow an easier analytical analysis [17]. A rst possible approach to this problem is to perform a non linear optimisation with constraints dened as: min(1 Kq Kq des + 2 des )
f

f [fmin ..fmax ]n

(20)

Following eq. (8), it is possible to modify the pretension forces, and consequently the mechanical joint stiffness, while ensuring that the joint torque is preserved. However, this is not valid when tendon force limits are reached. Indeed, neither the pseudo inversion nor the projection can guarantee that the desired tendon forces will be restricted to a given range. In the following, the objective is to build an algorithm to set the pretension forces to approximate the user required mechanical stiffness. Because of the contraints, this is not in general possible and only an approximative solution can be found. The presented algorithms must take into account the achievable force range and be implementable in realtime. As presented in several works [25], [26], the joint stiffness matrix can be obtained by the following transformation:

This optimisation does not ensure that the desired torque are achieved. However, in the usual application, the torque is more important than the mechanical stiffness and adding a new constraint in the problem solves this issue: (1 Kq Kq des ) min f (21) f [fmin ..fmax ]n P f = des

This latter formulation revealed to be complex to implement on the real-time machine, mainly due to the constraints. A reformulation of the problem (inspired by [25]), that ensures that the desired torques are achieved if possible, is proposed: min(1 Kq Kq des P ( q ) T |q=q0 = P (q )Kt (f )P (q )+ |q=q0 f Kq (q, f )|q=q0 = +2 1 (f , fmin ) f fmin q q (22) +3 2 (f , fmax ) f fmax ) (18) From eq.(18) and eq.(3), the following equation can be f = P + des + B, Rn dened: : [fmin ..fmax ] f
m

Rm Pf P T Kt (ft )P +

P (q ) q f

(19) The application is not bijective in general. For a given choice 70 60 K in [N/m] 50 40 30 20 10 0 10 20 30 40 50 60 70 80 90 Force in [N]
Fig. 3. Example of one tendon stiffness over force measurement

Where B = span(nullspace(P T )), i.e. B is a base of the nullspace of the coupling. Which ensures that P T B = 0 The weigthing, i > 0, i [1..3] are used to prevent an unreacheable coupling matrix from driving the search out of bounds. i (x, y ), i [1..2], (x, y ) R2 returns 0 if x > y otherwize y x. It is important to note that, in contrast to eq. (21) where the search is performed on f Rm , in eq. (21) Rn . This reduction of the search space provides a valuable run-time speed-up. Using this formulation, the particular shape of the nullspace of the coupling matrix is used advantageously to improve the search speed. The pseudo code corresponding to the search is: The projected gradient search algoritm: begin B = baseof N (P ) = 0 step = step0 costbest = inf for i := 1 to 50 do cost, grad = cost () if (cost < costbest )

of joint stiffness matrix and torque, there might only exist an approximative solution. Moreover, in the thumb of the hand

4218

then = grad step else step = step/2 end od where proc cost () f = P + des + B cost = costf (f ). end The key feature of the algoritm is to ensures that the desired torque is exactly achieved. Although it does not enforce that the force constraints are satied, it is in practice achieved since the boundary gains, 1 and 2 , are large which prevents the search from exceding the limits. In particular, this algorithm is extremely efcient with constant coupling matrices (i.e. the other ngers of the hand of the Hand Arm System). Indeed, if P is constant, the B matrix can be computed ofine. In case of time varying, or position varying coupling matrix, the main limitation of this algorithm is the need to compute online a singular value decomposition (or a pseudo inverse) online. Nonetheless, the special shape of the coupling matrix in the case of the thumb (block diagonal) allows for efcient implementation techniques. V. E FFECTIVE J OINT I MPEDANCE The effective impedance behaviour of the thumb is not equal to the one prescribed by the impedance controller. The general question of acheiving a complete link side impedance (often restricted to a second order dynamical system) is not treated here. However, a subproblem, which consist of improving the steady-state impedance (i.e. the stiffness), is considered. The mechanical stiffness and impedance stiffness have a serial connection (cf. g.4), and the effective stiffness K ef f is obtained as:

always results in an effective joint stiffness smaller than K imp . Given an effective stiffness K ef f,ref and a desired mechanical stiffness K mech,ref , the impedance controller can be improved by setting the controller stiffness accordingly to: 1 1 1 Kpimp = (K (24) ef f K mech ) Fig. 5 shows the results obtained using the modied impedance. The dashed black curve represents effective stiffness. The solid blue curve is the mechanical stiffness and the solid red curve is the controller stiffness. All quantities are model based, i.e. they are calculated based on the desired values and not the measured values (Fig. 5, top is a dynamic simulation and Fig. 5, bottom is a real hardware measurements). 3 2.5 K in [Nm/rad] 2 1.5 1 0.5 0 1.6 4.5 4 3.5 K in [Nm/rad] 3 2.5 2 1.5 1 Kef f 0.5 0 2 2.5 3 3.5 4 time in [s] 4.5 5 1.7 1.8 1.9 2 time in [s] 2.1 2.2

q Kmech Kimp

Fig. 4. The effective link stiffness is the serial combination of the controller and the mechanical (intrinsic) stiffness Fig. 5. Effective stiffness, mechanical stiffness and impedance controller stiffness (dotted black/bottom, solid blue/middle, solid red/top)
1 1 1 K ef f = (K mech + Kpimp )

(23)

It should be noted that the mechanical stiffness matrix can be singular if the mechanism is not tendon controllable [25]. Due to the serial interconnexion, the proposed controller

It is important to note that the resulting impedance stiffness might not be a positive denite matrix or can have very large elements. Therefore, some projection techniques might be required to nd the closest positive denite matrix

4219

[27]. In the realized implementation, the mechanical stiffness matrix is calculated from the desired values in order ensure passivity. VI. E XPERIMENTAL RESULTS AND SIMULATIONS In the previous sections a controller structure, as well as the joint position estimation algorithm were presented. This section reports the experimental results obtained with the real-time system2 . The implementation is done using MATLAB/Simulink on a QNX operating system. The search algorithms are evaluated, and the controller performance are presented along with simulations and experimental results. A. Search algorithms In order to evalute the algorithm a grid of tendon position vectors is generated from the kinematic model. The algorithm is evaluated on this vector grid and the resulting joint positions are compared to the ground truth. Fig. 6 depicts the results obtained with 30 steps. The two axis are representing the joint angles for the exion/extension q1 and abduction/adduction q2 motions. The red circles are the original points and the blue crosses are the estimated coordinates. The searches where always started from (0, 0), i.e. the green square. 1.5 1 0.5 q2 in [rad] 0 -0.5 -1 -1.5 -1.5

using the previous solution as a starting point the search always reaches the goal after only a few iterations.
1.5 1.5 1.5 1 1 1

0.5

0.5

0.5

0.5

0.5

0.5

1.5 1.5

0.5

0.5

1.5

1.5 1.5

0.5

0.5

1.5

1.5 1.5

0.5

0.5

1.5

(a)
1.5 1.5 1 1

(b)
1.5 1

(c)

0.5

0.5

0.5

0.5

0.5

0.5

1.5 1.5

0.5

0.5

1.5

1.5 1.5

0.5

0.5

1.5

1.5 1.5

0.5

0.5

1.5

(d)
1.5 1.5 1 1

(e)
1.5 1

(f)

0.5

0.5

0.5

0.5

0.5

0.5

1.5 1.5

0.5

0.5

1.5

1.5 1.5

0.5

0.5

1.5

1.5 1.5

0.5

0.5

1.5

(g)

(h)

(i)

Fig. 7. Results of the link position estimation. 7(a): 50 iterations, 0mm noise. 7(b): 50 iterations, 0.5mm noise. 7(c): 50 iterations, 1mm noise. 7(d): 30 iterations, 0mm noise. 7(e): 30 iterations, 0.5mm noise. 7(f): 30 iterations, 1mm noise. 7(g): 15 iterations, 0mm noise. 7(h): 15 iterations, 0.5mm noise. 7(i): 15 iterations, 1mm noise.

B. Controller performance The controller performance is presented by comparing the simulation and experiment results. The link trajectory obtained after a desired position step of 40 deg of the link is reported in Fig. 8. Because of the static friction, the two steady states (at 0 and at 0.7) are not reached. The 0.7 0.6 q in [rad] 0.5 0.4 0.3 0.2 0.1 0 2.4
Fig. 8.

-1

-0.5

0 0.5 q1 in [rad]

1.5

Fig. 6. Link position estimation: gradient search results with 30 iterations. The red circles are the original points and the blue crosses are the estimated coordinates.

To evaluate the robustness of the link side position estimation, the algorithm is given the vectors with a variable noise of 0mm, 0.5mm and 1mm to simulate the measurement inaccuracy. The number of steps is also varied in order to select the optimal value for the real-time code. Unlike the implemented code, the search is always started from q = [0, 0], which explains the bad results far from the origin. The results are reported in Fig. 7. The required number of iteration is easily achievable in realtime, moreover when
2 More details on the hardware communication infrastructure, using BiSS and Spacewire is found in [1].

2.6

2.8 3 3.2 time in [s]

3.4

Measured step response of 40 deg

measurement obtained with a load (at linear spring attached to a xed reference) at the thumb tip are presented in Fig. 9. Because of the load and the overall impedance the steady state position is, as expected, further away from the desired position. It is interesting to note that, at 0 deg, the mechanical stiffness is lower than in the deected position since the external spring load has no inuence when it is not elongated.

4220

As a result, some oscillations appear due to the reduced control bandwidth. 0.7 0.6 q in [rad] 0.5 0.4 0.3 0.2 0.1 0 6.6
Fig. 9.

6.8

7 7.2 time in [s]

7.4

7.6

Measured step response of 40 deg with a load at the thumb tip

VII. C ONCLUSION AND FUTURE WORK A controller structure for a tendon driven thumb has been presented along with simulation and experimental results. Since the link side is not measured, it was necessary to develop a method to estimate it. This expression cannot be obtained analytically and therefore a real-time optimization procedure is proposed The need for real-time implementation leads to a reformulation of the problem and an efcient search algorithm. The controller structure itself, a joint impedance controller with underlying tendon force controller, was proposed. Because the effective joint stiffness results of a serial interconnexion of impedances, the control stiffness gain was reshaped to improve the steadystate behavior. The results were asserted by confronting the experimental results with numerical simulations. The undergoing research concentrates on the development of an efcient method to inverse the system torque/stiffness equations under the tendon force contraints [26]. The method to share stiffness between the impedance controller stiffness and the mechanical stiffness, while satisfying the tendon forces range contraints is also part of the future work. ACKNOWLEDGEMENTS This work has been partially funded by the European Commissions Seventh Framework Programme as part of the project VIACTORS under grant no. 231554. R EFERENCES
[1] M. Grebenstein, A. Albu-Schaeffer, T. Bahls, M. Chalon, O. Eiberger, W. Friedl, R. Gruber, S. Haddadin, U. Hagn, R. Haslinger, H. Hoeppner, S. Joerg, M. Nickl, A. Nothelfer, F. Petit, J. Reill, N. Seitz, T. Wimboeck, S. Wolf, T. Wuesthoff, and G. Hirzinger, The DLR Hand Arm System, in International Conference on Robotics and Automation. IEEE, 2011. [2] M. Chalon, M. Grebenstein, and G. Hirzinger, The thumb : guidelines for a robotic design, in International Conference on Intelligent Robots and Systems. IEEE, 2010. [3] M. Grebenstein, M. Chalon, G. Hirzinger, and R. Siegwart, A method for hand kinematics designers 7 billion perfect hands, in Biomechanics, 2010.

[4] F. Laurin-Kovitz, K. Colgate, J. Carnes, and D. Steven, Design of components for programmable passive impedance, International Conference on Robotics and Automation, vol. 2, pp. 14761481, 1991. [5] K. Koganezawa and M. Yamazaki, Mechanical stiffness control of tendon-driven joints, in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, vol. 3, 2002, pp. 2280 2285. [6] H. Kino, N. Okamura, and S. Yabe, Basic characteristics of tendondriven manipulator using belt pulleys, in Proceedings of 2004 IEEE International Conference on Intelligent Robots and Systems, vol. 2, 2004, pp. 12871292. [7] R. Platt, M. Abdallah, C. Wampler, and B. Hargrave, Joint-space torque and stiffness control of tendon-driven manipulators, in International Conference on Intelligent Robots and Systems. IEEE, 2010. [8] G. Palli, Model and control of tendon actuated robots, Ph.D. dissertation, DEIS, University of Bologna, 2007. [9] S. Jacobsen, J. Wood, D. Knutti, and K. Biggers, Utah/mit dextrous hand - work in progress, in IJRR, vol. 3, no. 4, 1984, p. 2150. [10] J. Salisbury and M. T. Mason, Robot hands and the mechanics of manipulation, in MIT Press, 1985. [11] S. Jacobsen, H. Ko, E. Iversen, , and D. C., Antagonistic control of a tendon driven manipulator, in ICRA, 1989. [12] R. Schiavi, G. Grioli, S. Sen, and A. Bicchi, VSA-II: A novel prototype of variable stiffness actuator for safe and performing robots interacting with humans, in International Conference on Robotics and Automation, IEEE, Ed., 2008, pp. 21712176. [13] G.Palli, C.Melchiorri, T.Wimb ock, M.Grebenstein, and G. Hirzinger, Feedback linearization and simultaneous-position control of robot with antagonistic actuated joints, in International Conference on Robotics and Automation, 2007. [14] H. Kobayashi and R. Ozawa, Adaptive neural network control of tendon-driven mechanisms with elastic tendons, Automatica, vol. 39, pp. 15091519, 2003. [15] A. Albu-Sch affer, Ch. Ott, and G. Hirzinger, A unied passivity based control framework for position, torque and impedance control of exible joint robots, International Journal of Robotics Research, vol. 26, no. 1, pp. 23 39, January 2007. [16] , A unied passivity based control framework for position, torque and impedance control of exible joint robots, in International Symposium on Robotics Research, 2005. [17] T. Wimb ock, Ch. Ott, A. Albu-Sch affer, A. Kugi, and G. Hirzinger, Impedance control for variable stiffness mechanisms with nonlinear joint coupling, in International Conference on Intelligent Robots and Systems. IEEE, 2008, pp. 37963803. [18] L. Biagiotti, P. Tiezzi, C. Melchiorri, and G. Vassura, Modelling and controlling the compliance of a robotic hand with soft nger-pads, in Springer Tracts in Advanced Robotics, vol. 18, 2005, p. 55. [19] K. Koganezawa, Mechanical stiffness control for antagonistically driven joints, in IROS, 2005, p. 25122519. [20] K. Koganezawa, T. Nakazawa, and T. Inaba, Antagonistic control of multidof joint by using the actuator with non-linear elasticity, in ICRA, 2005, p. 22012207. [21] F. Werner, J. Reinecke, M. Chalon, and G. Hirzinger, Fas a exible antagonistic spring element for a high performance tendon driven hand, in submitted IROS, 2011. [22] T. N. Greville, Generalized Inverses. Springer-Verlag, 2003. [23] R. M. Murray, Z. Li, and S. S. Sastry, A Mathematical Introduction to Robotic Manipulation. CRC Press, 1994. [24] M. Grebenstein, M. Chalon, G. Hirzinger, and R. Siegwart, Antagonistically driven nger design for the anthropomorphic DLR Hand Arm System, in Humanoid. IEEE, 2010. [25] H. Kobayashi, K. Hyodo, and D. Ogane, On tendon-driven robotic mechanisms with redundant tendons, The International Journal of Robotics Research, vol. 17, no. 4, pp. 561571, 1998. [26] M. Chalon, T. Wimb ock, M. Grebenstein, and G. Hirzinger, Torque and workspace analysis for exible tendon driven mechanisms, in International Conference on Robotics and Automation. IEEE, 2010. [27] F. Petit, A. Albu-Schaeffer, and G. Hirzinger, Cartesian impedance control of a stiffness adjustable arm, in International Conference on Intelligent Robots and Systems. IEEE, 2011.

4221