Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Keywords: As industrial robots are applied in manufacturing industry on a large-scale and human intelligence is regarded
Physical human−robot interaction as an important part in manufacturing, physical human−robot interaction (pHRI) which integrates the strength
Dynamics and accuracy of robot with human operator’s ability of task cognition has drawn the attention of both academia
Quasi-static mode
and industry. However, an industrial robot without extra force/torque sensor for interacting force monitoring
Adaptive admittance control
cannot be used directly in pHRI, and research on pHRI of industrial robots remains a challenge. In this research,
a comprehensive dynamic model of an industrial robot in both dynamic mode and quasi-static mode is obtained to
calculate the external force produced by human operator in pHRI and enables sensorless pHRI for industrial robots
even in the environment with ambient vibration. Particularly, the dynamics in the process of mode switching
which has not been investigated by researchers is studied and compensated by an empirical but effective method.
Admittance control is used to transfer the detected force into reference position and velocity of the robot. RBF
(Radial Basis Function) network is used to update the damping parameter online in order to reduce the contact
force change and the contact force which makes pHRI more natural and easier. The stability of the controller is also
discussed. The proposed methods of external force detection and adaptive admittance control show satisfactory
behaviour in the experiments.
© 2017 Elsevier Ltd. All rights reserved.
∗
Corresponding author at: Department of Production Engineering, KTH Royal Institute of Technology, Stockholm 10044, Sweden.
E-mail address: bitao@kth.se (B. Yao).
https://doi.org/10.1016/j.rcim.2017.12.004
Received 28 April 2017; Received in revised form 21 November 2017; Accepted 21 December 2017
0736-5845/© 2017 Elsevier Ltd. All rights reserved.
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
detected based on the dynamic model even in an environment with [22] use convex optimisation method and prior knowledge of the exter-
ambient vibration. nal force to estimate the robot joint friction in velocities close to zero.
• On the basis of admittance control, we seek to update the damping This limits the application to environment where the prior knowledge
in order to reduce the contact force change and the contact force of the external force is unknown. Another method is to use a dithering
between the human operator and the robot and in this way pHRI can feedforward torque in the robot joint so that the joint friction is Coulomb
be more natural and easier. In this manner, the adaptive admittance friction [23−24]. The drawback of this method is that it is harmful to
control based on RBF network which is used to tune the parameter the mechanical structure of the robot. Although there are many studies
of the admittance controller adaptively is proposed. on the external force detection of the robots, the external force detec-
• One joint of an industrial robot is used to demonstrate the effec- tion of robots in the quasi-static mode is not studded adequately enough,
tiveness of the proposed methods for external force detection of the particularly, the method of mode switching and the dynamic process of
robot in both dynamic mode and quasi-static mode and the adaptive mode switching between the dynamic mode and the quasi-static mode.
control of the robot. The methods can be extended to an n-DOF robot
without any theoretical barrier.
2.2. Control strategies in pHRI
The rest of this paper is organised as follows. Section 2 summarises
the state-of-the-art in external force detection and control strategies The impedance/admittance control strategy realises a compliance
of pHRI, and highlights the novelty and innovation of our method. control and is most popular in pHRI. Systems which accept efforts to pro-
Section 3 deals with the robotic modelling, including the Stribeck fric- duce flows are called admittances, while systems which accept flows to
tion modelling, the inertial parameter estimation, the dynamic mod- produce efforts are called impedances [25]. In [26], different impedance
elling of the robot in quasi-static mode, and the method of mode switch- modulation strategies are studied to modify the impedance of a redun-
ing. In Section 4, an adaptive admittance control strategy based on RBF dant manipulator for pHRI. The redundancy is made use of to make the
network is proposed. The control stability is also discussed in this sec- robotic equivalent inertia at the end-effector close to the desired iner-
tion. Section 5 provides the experimental results of sensorless external tia. The control characteristics of a human operator who interacts with a
force detection and adaptive admittance control of a robot in pHRI. The human−robotic system using an impedance controller are investigated
conclusions and future work are discussed in Section 6. in [27]. Other examples of applying impedance control in pHRI can be
found in [28–30], etc.
2. Related work Variable impedance/admittance control has been adopted to achieve
various goals. In [31], the manipulator end-effector force is used to ad-
In this section, we first summarise the state-of-the-art in external just the admittance parameters and to reduce the oscillations in pHRI.
force detection of the robot, and then review the control strategies in Model-free continuous critic learning is proposed in [32] to optimise the
pHRI. impedance model and realise trajectory tracking and force regulation in
HRI, but experimental verification is not provided. In [33–36], Neural
2.1. External force detection of the robot network in the inner loop of a two loop controller is used to learn the dy-
namic model of the robot and make the robot response like a prescribed
pHRI is getting more and more attention from both academic and in- impedance, and the outer loop takes into account the human dynamics
dustrial sectors. The detection of external force produced by the human and adapt the robot impedance model, so that the robot system assists
operator is a prerequisite. Proprioceptive sensation is a direct method to humans with varying levels of skill to achieve task-specific objectives.
measure the external force. It includes haptic devices on the robot such However, the dynamic model of the robot can be complex with coupled
as tactile sensor skin [6] and force/torque sensors at the end-effector of dynamics of each joint and this limits the accuracy of the learned dy-
the robot [7,8]. In [3,4], a human operator steers a robot with the help of namic model of the robot. The human dynamics is also considered in the
force sensor that measures the user’s applied force. Some light-weighted adaptive impedance control in pHRI and modelled as a first-order lag
collaborative robots which use joint torque sensors to detect the force in [37]. Variable impedance is used in [39] to compensate for the un-
produced by human operators have been developed and introduced to modelled uncertainties of the learned task model during pHRI. Variable
the robot market, such as Kuka IIWA [9], ABB YuMi [10], Kinova [11], impedance which changes with the stiffness of the human arm can also
RethinkRobotics [12], etc. They attract much attention of researchers, be used to improve the stability of human−robot cooperative task [38].
for example, in a pHRI based homokinetic joint assembly case in [13], Besides compliance, there are other metrics to characterise the con-
the human interacts with Kuka IIWA. The performance of sensor-based trol strategies of pHRI. Positioning accuracy is one of the metrics. Hid-
solution depends on the accuracy of the sensor. Usually, the higher the den Markov Models are used in automatic segmentation and recogni-
accuracy of the sensor is, the heavier the cost is. Moreover, force/torque tion of user motions to adjust the stiffness of virtual fixtures for human-
sensors add weight to the robot. machine cooperative systems [40,41]. In this way, the accuracy of curve
Besides sensor based methods, some researchers have studied sen- following and object avoidance in robot-assisted manipulation can be
sorless methods of external force detection. Disturbance observers have improved. The movement of human hand follows some characteristics
been used in the external force detection of robots, such as [14−17]. that can be used as the metrics of pHRI. Minimum hand jerk and joint
The authors of [18] propose a filter-based method for the detection of torque change are the characteristics of unconstrained point-to-point
external force on an industrial robot and the robot can react to human’s movements of human hand [42]. In pHRI, the movement of human hand
push/pull. In these studies [14–18], the output of the servo level robot is constrained by robots or the object that is manipulated by human op-
controller is needed, but this is not available for most industrial robots erator and the robot simultaneously. In the constrained environment,
whose controllers are not open to robot users. Generalised momentum combined minimum hand contact force change and the actuating force
is proposed by [19] and can be used to monitor the external force on (torque or muscle force) change characterise the hand movement [43].
the robot, such as in [20,21]. However, the inertia parameters in the There are some other metrics, such as time-to-completion [44].
generalised momentum are needed. There are other issues in the control of pHRI such as lead-follow role
In the dynamic mode, the robot joint friction is usually deemed as assignment [45,46], safety issue [47,48], etc. This work mainly focuses
Coulomb friction. However, in the quasi-static mode, the joint friction is on the adaptive admittance control of the robot to make pHRI easier
quite different from Coulomb friction and is complex. In order to facili- and more natural by reducing the human−robot contact force and the
tate the external force detection in the quasi-static mode, the authors of contact force change.
159
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
Fig. 1. KUKA KR 6 R700 sixx. Fig. 3. Excitation trajectory for the extraction of Stribeck friction with a velocity of
0.312 rad/s.
where FC is the Column friction, FS the static friction, v the velocity of the
Fig. 2. Scheme of KUKA robot controller. joint, vS the Stribeck velocity, 𝛿 S an additional empirical parameter, and
Fv the viscous friction. It is common to accept 𝛿 S as a constant between
0.5 and 2 [53].
3. Dynamic modelling of robot The friction of Joint 3 is extracted by constant speed experiments in
which the torque caused by acceleration is zero. Joint 3 is commanded
In this research, KUKA KR 6 R700 sixx equipped with a KR C4 con- to move at constant speed clockwise and counterclockwise, respectively.
troller and KSS 8.3 software system is used to perform the modelling An example of excitation trajectory for the extraction of Stribeck friction
and control of the industrial robot in pHRI. Joint 3 of the robot is used is shown in Fig. 3.
to demonstrate the effectiveness of the proposed methods. These meth- When Joint 3 moves clockwise and other joints keep stationary, the
ods can be extended to other joints of the robot. To achieve this, the torque of joint 3 is
additional work is to identify the inertial parameters of the robot by
Γ+ = Γf + 𝑄(𝑞 ) (2)
stimulating all joints of the robot simultaneously, which can be found
in [49]. The coordinate of Joint 3 is shown in Fig. 1. where Γf is the joint friction torque, q is the joint position, and Q(q) is
The scheme of KUKA Robot Controller (KRC) is shown in Fig. 2 [50]. the joint torque caused by the gravity of links 3, 4, 5, and 6 (link i is the
The HMI (Human Machine Interface) communicates with the KRC in ev- link that connects joint i with joint i + 1) as shown in Fig. 1.
ery 12 ms. We use a SUB file in KRC which works in parallel with the When Joint 3 moves counter-clockwise and other joints keep station-
robot controller and is processed by the controller interpreter to obtain ary, the torque of joint is
the value of system variables $AXIS_ACT_MEAS (measured joint posi- Γ − = − Γ f + 𝑄 (𝑞 ) (3)
tions) and $TORQUE_AXIS_ACT (measured output joint torques). The
sampling rate is approximately 83 Hz and is found to be sufficient for From Eqs. (2) and (3), it can be derived that
the dynamic modelling. A three order low-pass IIR Butterworth filter is Γ+ − Γ −
Γf = (4)
used to filter the noise of the measured torque as well as the velocity and 2
acceleration of the joint which are obtained by differential operation on Based on Eq. (4), the joint friction can be extracted from the mea-
the joint positions. The cut-off frequency of the filter is set as 5 Hz as sured joint torques, as shown in Fig. 4.
the human hand movements do not exceed 5–10 Hz [51]. The extracted joint friction and optimisation approach is used to es-
timate the unknown parameters in model (1). The objective function of
3.1. Stribeck friction of robot joint the optimisation is given by
𝑗=𝑘 [
∑ ( )]2
In the dynamic mode, the friction between two contacted surfaces is ( ) 𝛿S
min𝑓 = Γf − 𝐹C + 𝐹S − 𝐹C exp−|𝑣∕𝑣S | + 𝐹v 𝑣 (5)
nonlinear with the velocity of the surfaces’ relative movement. Stribeck 𝑗=1
160
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
Table 1
Estimated values of Stribeck parameters.
FC FS vS 𝛿S Fv
161
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
𝑊 𝐹𝑖 = 𝑘𝑖 𝑧 − 𝜍𝑖
if||𝑧 − 𝜍𝑖 || < 𝑖 , then
𝑘𝑖 𝜍 = const
( ) 𝑖
⎧𝐹𝑖 = sgn 𝑧 − 𝜍𝑖 𝑊𝑖 (9)
⎪
else⎨ ( ) 𝑊𝑖
⎪𝜍𝑖 = 𝑧 − sgn 𝑧 − 𝜍𝑖 𝑘
⎩ 𝑖
162
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
Table 2 the friction should be constant), the joint torque also changes obviously
Parameters of three Maxwell slip elements in parallel.
(from 4.83 Nm to 7.24 Nm). This phenomenon may be attributed to the
Element 1 Value Element 2 Value Element 3 Value unmodelled dynamics in the reducer in the joint, and will cause error in
k1 35,894 k2 160,065 k3 51,893
the predicted joint torque. The error will be examined by experiments
W1 7.3406 W2 1.371 W3 6.288 in Section 5.1 and compensated by an empirical but effective method.
𝜍1 −0.00041 𝜍2 −0.00048 𝜍3 0.00027 Therefore, without external force, the dynamics of Joint 3 in quasi-
static mode can be modelled as
𝑀X cos(𝑞 ) + 𝐹p + Γu = Γ (11)
Models (6) and (11) can be applied to dynamic mode and quasi-
static mode, respectively. A criterion should be adopted to guide the
selection of model to calculate the predicted joint torque. As the two
models apply to different regimes of friction, we define a threshold 𝑞̇ thr
of absolute joint angular velocity to determine the mode of joint move-
ment and guide the model selection. If the absolute joint angular veloc-
ity is smaller than 𝑞̇ thr , it is considered that the joint is in quasi-static
mode and model (6) will be adopted to calculate the predicted torque.
Otherwise, model (11) will be used to calculate the predicted torque.
However, there are some challenges to choose an appropriate 𝑞̇ thr . If 𝑞̇ thr
is too small (the extreme is zero), the movement which is actually in
the pre-sliding regime may be missed and this will bring difficulties to
identifying the initial value of 𝜍 i as discussed in the following.
According to model (9), the initial value of 𝜍 i is critical to the calcula-
tion of the predicted friction of joint in pre-sliding regime. Although the
initial value of 𝜍 i can be estimated based on optimisation approach, it
needs to be estimated again if the online monitoring of the joint position
was interrupted. Therefore, a practical and efficient method is needed.
Suppose a state when the asperity deflection of the contacting surfaces
reaches the maximum and the output force of each Maxwell slip ele-
ment becomes saturated. This state is chosen as the initial state for the
obtaining of 𝜍 i . In this state, 𝜍 i can be obtained as
𝑊i
𝜍i = sgn(𝑣) (12)
𝑘i
If there is a relative movement between the contacting surfaces after
reaching this initial state, 𝜍 i is updated based on model (9). The common
displacement z is calculated as the position increment relative to the
joint position when the joint reaches the initial state for the first time.
Fig. 10. Results of the step position change experiment: (a) position of Joint 3, (b) first According to the above discussion, 𝑞̇ thr should be big enough so that
difference of position in (a) and the enlarged figure of the first difference of position in the initial state which is critical to the obtaining of 𝜍 i can be included in
the dashed box, (c) torque of Joint 3. the quasi-static mode. However, if 𝑞̇ thr is too big, there may be obvious
viscous friction in the joint and the friction calculated by model (11) may
not be accurate because viscous friction is not included in it. With these
a target position and then stops. Fig. 10 shows the data in the test. It can considerations in mind, 𝑞̇ thr is set as 0.0017 rad/s in this research, the
be seen from Fig. 10(a) that the friction is sliding friction before point corresponding angular velocity in degrees is about 0.1°/s.
A since the joint position decreases from −0.2105 rad to −0.2123 rad The initial value of 𝜍 i can be obtained by moving the robot joint
monotonously in 0.24 s (about 0.43°/s). Consequently, before point A, slightly as shown by the green line in Fig. 11. After this movement, the
the asperity of contacting surfaces of Joint 3 keeps at the maximum. joint stops and is in quasi-static mode for the first time and the asperity
From Fig. 10(b), we can see that the first difference of joint position deflection of the contacting surfaces of the joint reaches the maximum.
keeps decreasing from point A to point C and it means that the asperity The initial value of 𝜍 i is, therefore, can be calculated by (12). We can call
of the contacting surfaces still keeps at the maximum. Considering that this process the initiation stage. Then, the joint goes into dynamic mode
the velocity of Joint 3 from point A to point C is near zero, the joint fric- if there is detected external force. The positions where the robot joint
tion keeps at the maximum pre-sliding friction and can be considered goes from dynamic mode to quasi-static mode can then be captured by
as a constant (equal to the total output of all the Maxwell slip elements) comparing the measured absolute angular velocity with 𝑞̇ thr . Based on
during this process. Meanwhile, the position changes only slightly from this method, we can switch between the dynamic model and the quasi-
point A to point C, and the gravity-related dynamics on Joint 3 is nearly static model.
constant. The acceleration is also near zero from point A to point C and
thus the torque due to the moment of inertia is almost zero. It seems to 3.5. External force detection
come to a conclusion that the joint torque from point A to point C was
constant. However, the joint torque changes from 2.79 Nm to 7.24 Nm Models (6) and (11) are used to fully characterise the dynamics of
when Joint 3 goes from point A to point C. Besides, while the first differ- the robot and to calculate the predicted joint torque without consider-
ence of position is about zero from point B to point C (which means that ing the external force. Specifically, model (6) is used to characterise the
163
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
164
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
The centres are important and are set according to the force or force
change level in the experiment. If the centres are set to too low or too
high, updating the weights of radial basis functions will be poor.
5. Experiments
165
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
is determined by the accuracy of the dynamic model which is affected by quasi-static mode. As shown in Fig. 15(d), Joint 3 successfully switches
the method of parameter identification of unknown parameters in (11), between the quasi-static mode and the dynamic mode. However, there
the unmodelled robotic dynamics, and the high noise in the feedback are some phenomena that are noteworthy. There is a time delay of robot
joint torque. In the future work, we will perform a Benchmark test to movement when there is an external force at about 9.8 s and 18.6 s. This
further evaluate the method in this work. The resolution of the detected delay is longer than the cycle of control which is about 24 ms. This is
force depends on the sensor that measures feedback joint torque. because all joints of the robot are synchronised in a way that all joints
start and stop moving at the same time when the motion control com-
5.2. Experiment 2: switching between dynamic and quasi-static modes mand PTP (point − to − point motion) of KRL (Kuka Robot Language) is
executed. Therefore, Joint 3 may move very slowly if the trajectory of
Similar to Experiment 1, Joint 6 is commanded to move randomly to joint 6 is much longer than joint 3′s. Another phenomenon is that there
simulate the ambient vibration during the experiment and other joints are some peaks in the torque residual in Fig. 15(d). These peaks are at
do not move. The initial position of Maxwell slip elements of joint 3 the position where the joint changes from quasi-static mode to dynamic
𝜍 i is initiated by a movement as shown by the green line of Fig. 11. mode. This is caused by the tiny mechanical backslap in the reducer
𝑞̇ thr is set as 0.0017 rad/s. During the experiment, the human operator which is difficult to model.
applies a force near the end-effector for two times to produce a torque
about the z axis of Joint 3 (as shown in Fig. 1) clockwise and coun-
terclockwise respectively, each for one time. The robot is commanded 5.3. Experiment 3: admittance control with constant damping
to move to a reference position proportionally to the torque residual
with a constant velocity. Fig. 15(d) shows the torque residual that is In this experiment, the robot is controlled under the contact force
calculated based on model (13). The quasi-static mode is shown by the between the human operator and the robot. Dynamic models (6) and
yellow areas in Fig. 15(d) while the dynamic mode is shown by the pink (11) are applied for external force detection on the robot. The admit-
areas. It shows the process of switching between dynamic mode and tance control is applied with a constant damping of 50 Nm s/rad and
quasi-static mode. First, Joint 3 goes into quasi-static mode after the a constant moment of inertia 20 kg/m2 . The results are obtained and
initiation stage. At the time about 10 s, there is an external force which shown in Fig. 16. Fig. 16(a) and (c) show that the position of Joint 3
is produced by the human operator and higher than the threshold (5 can be changed frequently under the external force of the human oper-
Nm) shown by the blue dash line in Fig. 15(d). Consequently, the posi- ator. During t1 −t2 , the external torque on joint 3 is about 10 Nm, the
tion of Joint 3 is commanded to increase proportionally to the torque joint angular velocity keeps approximately constant at 0.2 rad/s and the
residual by the program in the frontend computer and Joint 3 goes into acceleration of the joint can be regarded as approximately 0. According
dynamic mode. The movement of Joint 3 is shown in Fig. 15(a). After to (15), the resisting force of the admittance controller is about 10 Nm
the human operator stops acting a torque on Joint 3 and the last com- (50 Nm·s/rad × 0.2 rad/s) and this is nearly equal to the external force.
mand is executed, Joint 3 stops. During the process that Joint 3 goes This verifies the validity of the admittance controller. However, there
from dynamic mode to quasi-static mode, the point where the absolute are some spikes with high torque residuals when the joint starts to move
joint velocity is 𝑞̇ thr will be captured and the position at this point is and stops. This is due to the mechanical backslash that is hard to model
used as a new initial position for the calculation of predicted torque in and not included in the model (11).
166
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
6. Conclusions
20
torque even in a vibration environment. The torque residuals reflect the
(Nm)
(c) 10
contact force between the robot and the human operator without using
0
extra sensors. The torque residuals are then transferred into reference
0 1 2 3 4 5 6 7 8
position and velocity of the robot by admittance control. The damping
Torque residual
2
change (Nm)
(d) 0
of admittance control is updated online in pHRI to reduce the contact
force change and the contact force and thus make pHRI easier and more
-2
0 1 2 3 4 5 6 7 8 natural. The experimental results show the effectiveness of the proposed
4
dynamic model and the adaptive control method. This makes the indus-
Cost
(e) 2 trial robot a possible partner in pHRI. However, there are some problems
0 that can be improved in our future work, such as the friction change
0 1 2 3 4 5 6 7 8
when the joint temperature changes. Moreover, the safety issue in pHRI
Time (s)
is also important. The standard on collaborative robots such as ISO/TS
Fig. 17. Adaptive admittance control: green lines refer to the results of adaptive admit- 15,066:2016 can be incorporated in the future.
tance control while blue lines refer to the results of admittance control with constant
damping. (For interpretation of the references to colour in this figure legend, the reader
is referred to the web version of this article.) Acknowledgement
167
B. Yao et al. Robotics and Computer–Integrated Manufacturing 51 (2018) 158–168
[8] L. Rozo, S. Calinon, D.G. Caldwell, P. Jiménez, C. Torras, Learning physical collab- [34] I. Ranatunga, S. Cremer, D.O. Popa, F.L. Lewis, Intent aware adaptive admittance
orative robot behaviors from human demonstrations, IEEE Trans. Robot. 32 (2016) control for physical Human-Robot Interaction, in: Proceedings of the IEEE Interna-
513–527. tional Conference on Robotics and Automation, 2015, pp. 5635–5640.
[9] S. Shepherd, A. Buchstab, Kuka robots on-site, in: Robotic Fabrication in Architec- [35] I. Ranatunga, F.L. Lewis, D.O. Popa, S.M. Tousif, Adaptive admittance control for
ture, Art and Design, 2014, Springer, 2014, pp. 373–380. human−robot interaction using model reference design and adaptive inverse filter-
[10] ABB Group, IRB 14000 YUMI, 2015 (accessed 17.02.10). http://new. ing, IEEE Trans. Control Syst. Technol. 25 (2017) 278–285.
abb.com/products/robotics/industrial-robots/yumi . [36] H. Modares, I. Ranatunga, F.L. Lewis, D.O. Popa, Optimized assistive human−robot
[11] Kinova robotics, FEATURES, in, http://www.kinovarobotics.com/innovation- interaction using reinforcement learning, IEEE Trans. Cybern. 46 (2016) 655–667.
robotics/products/robot-arms/, (accessed 17.11.15). [37] S. Suzuki, K. Furuta, Adaptive impedance control to enhance human skill on a haptic
[12] RethinkRobotics, Sawyer, in, http://www.rethinkrobotics.com/sawyer/, (accessed interface system, J. Control Sci. Eng. 2012 (2012) 1–10.
17.11.15). [38] T. Tsumugiwa, R. Yokogawa, K. Hara, Variable impedance control based on esti-
[13] A. Cherubini, R. Passama, A. Crosnier, A. Lasnier, P. Fraisse, Collaborative manu- mation of human arm stiffness for human−robot cooperative calligraphic task, in:
facturing with physical human–robot interaction, Robot. Comput.-Integr. Manuf. 40 Proceedings of the 2002 IEEE International Conference on Robotics and Automation
(2016) 1–13. (Cat. No.02CH37292), 641, 2002, pp. 644–650.
[14] S. Katsura, Y. Matsumoto, K. Ohnishi, Modeling of force sensing and validation of [39] E. Gribovskaya, A. Kheddar, A. Billard, Motion learning and adaptive impedance for
disturbance observer for force control, IEEE Trans. Ind. Electron. 54 (2007) 530–538. robot control during physical interaction with humans, in: Proceedings of the 2011
[15] L. Chan, F. Naghdy, D. Stirling, Extended active observer for force estimation IEEE International Conference on Robotics and Automation, 2011, pp. 4326–4332.
and disturbance rejection of robotic manipulators, Robot. Auton. Syst. 61 (2013) [40] M. Li, A.M. Okamura, Recognition of operator motions for real-time assistance using
1277–1287. virtual fixtures, in: Proceedings of the IEEE11th Symposium on Haptic Interfaces
[16] N. Surapong, C. Mitsantisuk, Position and force control of the SCARA robot based for Virtual Environment and Teleoperator Systems, 2003. HAPTICS 2003, 2003,
on disturbance observer, Proc. Comput. Sci. 86 (2016) 116–119. pp. 125–131.
[17] N. Shimada, T. Yoshioka, K. Ohishi, T. Miyazaki, Y. Yokokura, Variable dynamic [41] J.J. Abbott, P. Marayong, A.M. Okamura, Haptic Virtual Fixtures for Robot-Assisted
threshold of jerk signal for contact detection in industrial robots without force sen- Manipulation, in: S. Thrun, R. Brooks, H. Durrant-Whyte (Eds.), Proceedings of the
sor, Electr. Eng. Jpn. 193 (2015) 43–54 (English translation of Denki Gakkai Ron- Results of the 12th International Symposium on Robotics Research: ISRR, Berlin,
bunshi). Heidelberg, Springer, 2007, pp. 49–64.
[18] M. Geravand, F. Flacco, A. De Luca, Human-robot physical interaction and collab- [42] M. Svinin, K. Ohta, Z.W. Luo, S. Hosoe, Towards understanding of human movements
oration using an industrial robot with a closed control architecture, in: Proceed- constrained by the external environment, in: Proceedings of the 2003 IEEE/RSJ
ings of IEEE International Conference on Robotics and Automation (ICRA), 2013, International Conference on Intelligent Robots and Systems (IROS 2003) (Cat.
pp. 4000–4007. No.03CH37453), 151, 2003, pp. 155–161.
[19] A.d. Luca, R. Mattone, Sensorless robot collision detection and hybrid force/motion [43] K. Ohta, M.M. Svinin, Z. Luo, S. Hosoe, R. Laboissière, Optimal trajectory formation
control, in: Proceedings of the 2005 IEEE International Conference on Robotics and of constrained human arm reaching movements, Biol. Cybern. 91 (2004) 23–36.
Automation, 2005, pp. 999–1004. [44] A. Steinfeld, T. Fong, D. Kaber, M. Lewis, J. Scholtz, A. Schultz, M. Goodrich,
[20] M. Ragaglia, A.M. Zanchettin, L. Bascetta, P. Rocco, Accurate sensorless lead- Common metrics for human−robot interaction, in: Proceedings of the 1st ACM
-through programming for lightweight robots in structured environments, Robot. SIGCHI/SIGART conference on Human-robot interaction, ACM, 2006, pp. 33–40.
Comput.-Integr. Manuf. 39 (2016) 9–21. [45] A. Mörtl, M. Lawitzky, A. Kucukyilmaz, M. Sezgin, C. Basdogan, S. Hirche, The role
[21] Z. Liu, F. Yu, L. Zhang, T. Li, Real-time estimation of sensorless planar robot contact of roles: physical cooperation between humans and robots, Int. J. Robot. Res. 31
information, J. Robot. Mechatron. 29 (2017) 557–565. (2012) 1656–1674.
[22] M. Linderoth, A. Stolt, A. Robertsson, R. Johansson, Robotic force estimation using [46] Y. Li, K.P. Tee, R. Yan, W.L. Chan, Y. Wu, D.K. Limbu, Adaptive optimal control
motor torques and modeling of low velocity friction disturbances, in: Proceedings of for coordination in physical human−robot interaction, in: Proceedings of the IEEE
2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2013, International Conference on Intelligent Robots and Systems, 2015, pp. 20–25.
pp. 3550–3556. [47] S. Norouzzadeh, T. Lorenz, S. Hirche, Towards safe physical human−robot interac-
[23] A. Stolt, A. Robertsson, R. Johansson, Robotic force estimation using dithering to tion: an online optimal control scheme, in: Proceedings of the IEEE International
decrease the low velocity friction uncertainties, in: Proceedings of 2015 IEEE Inter- Workshop on Robot and Human Interactive Communication, 2012, pp. 503–508.
national Conference on Robotics and Automation (ICRA), 2015, pp. 3896–3902. [48] M. Kimmel, M. Lawitzky, S. Hirche, 6D workspace constraints for physical
[24] H. Cho, M. Kim, H. Lim, D. Kim, Cartesian sensor-less force control for industrial human−robot interaction using invariance control with chattering reduction, in:
robots, in: Proceedings of IEEE/RSJ International Conference on Intelligent Robots Proceedings of the IEEE International Conference on Intelligent Robots and Systems,
and Systems, 2014, pp. 4497–4502. 2012, pp. 3377–3383.
[25] G. Ferretti, G. Magnani, P. Rocco, Impedance control for elastic joints industrial [49] W. Khalil, E. Dombre, Modeling, Identification and Control of Robots, Butter-
manipulators, IEEE Trans. Robot. Autom. 20 (2004) 488–498. worth-Heinemann, 2004.
[26] F. Ficuciello, L. Villani, B. Siciliano, Variable impedance control of redundant manip- [50] E. Prassler, R. Bischoff, W. Burgard, R. Haschke, M. Hägele, G. Lawitzky, B. Nebel,
ulators for intuitive human robot physical interaction, IEEE Trans. Robot. 31 (2015) P. Plöger, U. Reiser, M. Zöllner, Towards Service Robots For Everyday environments:
850–863. Recent Advances in Designing Service Robots For Complex Tasks in Everyday Envi-
[27] T. Tsuji, Y. Tanaka, Tracking control properties of human−robotic systems based on ronments, Springer, 2012.
impedance control, IEEE Trans. Syst. Man Cybern. - Part A: Syst. Hum. 35 (2005) [51] F. Parietti, G. Baud-Bovy, E. Gatti, R. Riener, L. Guzzella, H. Vallery, Series viscoelas-
523–535. tic actuators can match human force perception, IEEE/ASME Trans. Mechatron. 16
[28] Y. Li, S.S. Ge, Human-robot collaboration based on motion intention estimation, (2011) 853–860.
IEEE/ASME Trans. Mechatron. 19 (2014) 1007–1014. [52] H. Olsson, K.J. Åström, C. Canudas de Wit, M. Gäfvert, P. Lischinsky, Friction models
[29] P.D. Labrecque, J.M. Haché, M. Abdallah, C. Gosselin, Low-impedance physical and friction compensation, Eur. J. Control 4 (1998) 176–195.
human−robot interaction using an active-passive dynamics decoupling, IEEE Robot. [53] A.C. Bittencourt, S. Gunnarsson, Static friction in a robot joint—modeling and iden-
Autom. Lett. 1 (2016) 938–945. tification of load and temperature effects, J. Dyn. Syst. Meas. Control 134 (2012)
[30] E. Magrini, F. Flacco, A. De Luca, Control of generalized contact motion and force in 051013.
physical human−robot interaction, in: Proceedings of IEEE International Conference [54] V. Lampaert, J. Swevers, F. Al-Bender, Modification of the Leuven integrated friction
on Robotics and Automation, 2015, pp. 2298–2304. model structure„ IEEE Trans. Autom. Control 47 (2002) 683–687.
[31] V. Okunev, T. Nierhoff, S. Hirche, Human-preference-based control design: adap- [55] KUKA. System Software, in, https://www.kuka.com/en-se/products/robotics-
tive robot admittance control for physical human−robot interaction, in: Proceedings systems/software/system-software/kuka_systemsoftware, (accessed 17.11.15).
of IEEE International Workshop on Robot and Human Interactive Communication, [56] F. Sanfilippo, L.I. Hatledal, Z. Houxiang, M. Fago, K.Y. Pettersen, Controlling Kuka
2012, pp. 443–448. industrial robots: flexible communication interface JOpenShowVar„ IEEE Robot. Au-
[32] C. Wang, Y. Li, S.S. Ge, K.P. Tee, T.H. Lee, Continuous critic learning for robot control tom. Mag. 22 (2015) 96–109.
in physical human−robot interaction, in: Proceedings of International Conference on
Control, Automation and Systems, 2013, pp. 833–838.
[33] B. Alqaudi, H. Modares, I. Ranatunga, S.M. Tousif, F.L. Lewis, D.O. Popa, Model
reference adaptive impedance control for physical human−robot interaction, Control
Theory Technol. 14 (2016) 68–82.
168