2022 25th International Conference on Mechatronics Technology (ICMT)

Force Observer-Based Admittance Control Design

for Robot Manipulators
Hoang Vu Dao Kyung Sin Kwak Kyoung Kwan Ahn
Graduate School of Mechanical Graduate School of Mechanical School of Mechanical Engineering
Engineering Engineering University of Ulsan
2022 25th International Conference on Mechatronics Technology (ICMT) | 978-1-6654-6195-5/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICMT56556.2022.9997788

University of Ulsan University of Ulsan Ulsan, South Korea

Ulsan, South Korea Ulsan, South Korea [email protected]
[email protected] [email protected]

Abstract— This paper proposed a novel admittance quite high, which may deteriorate the transient response of the
controller for robot manipulators suffering from nonlinear control system.
dynamics and force disturbances without contact force sensors.
A nonlinear observer is designed to simultaneously estimate To improve the performance of the admittance control
both contact force and unmeasurable joint velocities. These system, in our previous studies [4], the extended sliding mode
estimated values are fed back to the model-based admittance observer (ESMO) has been proposed which relies on the
control framework which includes the outer admittance control domination of the “sign” function compared to the
loop and inner position control loop. Simulation with a two- proportional term of the estimation errors. However, the order
degree-of-freedom (2-DOF) manipulator is conducted to verify of the ESMO is the same as the ESO which remains an open
the effectiveness of the proposed algorithm. problem for the design of high-accuracy force estimators.
Keywords— admittance control, force sensorless, robot Based on the above observation, in this paper, a nonlinear
manipulator. observer is designed to estimate unknown contact force and
joint velocities for the admittance control purpose of robot
I. INTRODUCTION manipulators. Compared to other force observer techniques,
Robot manipulators are widely utilized in various the proposed observer reduces the computational burden and
industrial applications such as robotic assembly [1], material provides a fast response when contact motion occurs. After
handling [2], automated welding [3], etc. due to the ability to estimation, the observed force and velocities are fed back into
quickly and accurately complete repetitive tasks. To produce the admittance control framework which guarantees the
this advanced performance, the control tasks are classified into admittance behavior of the control system. The stability of the
position control tasks and force control tasks which regulate proposed system is guaranteed by the Lyapunov theory.
the behavior of the robot in free motion and constraint motion, Finally, different simulation scenarios with a 2-DOF robot
respectively. However, for applications including both kinds manipulator are investigated to validate the performance of
of motion during operation, it is impossible to guarantee the proposed algorithm.
control performance by only using a position controller or The paper is organized as follows: Section II described the
force controller [4]. system modeling. Section III presents the proposed algorithm.
To overcome this problem, admittance control was Numerical simulation is discussed in Section IV. Finally, the
proposed which regulates the relationship between the contact conclusion is given in Section V.
force and the end-effector position [5]. In free motion, it II. SYSTEM MODELING
behaves like a normal position control when the contact force
vanishes. In contrast, when the contact motion happens, it The manipulator dynamics are expressed as
regulates the position of the end-effector depending on the
M (q )q&& + C (q, q& )q& + G (q ) = u + J (q )T Fc (1)
designed “stiffness” of the system. In order to reduce the
hardware complexities for the design of admittance control,
where q ∈ R n , M (q ) ∈ R n×n , C (q, q& ) , and G (q ) are the
force-sensorless approaches were proposed which utilize
force observers to estimate the contact force instead of force vector of joint angles, inertia matrix, Coriolis and centrifugal
sensors [6]. In [7], the generalized momentum observer matrix, and gravity matrix, respectively. u is the control
(GMO) was proposed to observe the interaction force based force/torque, and Fc is the interaction force/moment between
on the position and velocity information. However, since the the end-effector of the robot and the environment. J (q) is the
velocity information is disturbed by noise and requires vector of lumped disturbances and uncertainties existing in the
velocity sensor installment, this method is quite limited. system, which is computed by
Another approach called extended state observer (ESO) can
be considered which simultaneously estimates both lumped ∂p
disturbances and unmeasurable joint velocity by utilizing only J (q ) = (2)
∂q
position measurement [8]. However, the order of the ESO is
where p is the position of the end-effector. For convenience,
This research was supported by Basic Science Program through the the system dynamics are transformed into
National Research Foundation of Korea (NRF) funded by the Ministry of
Science and ICT, South Korea (NRF 2020R1A2B5B03001480) and x&1 = x2
“Regional Innovation Strategy (RIS)” through the National Research (3)
Foundation of Korea (NRF) funded by the Ministry of Education (MOE) x& 2 = g ( x1 )u + f ( x1 , x2 ) + d
(2021RIS-003).

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 01,2023 at 11:31:31 UTC from IEEE Xplore. Restrictions apply.
 where x1 = q , x2 = q& , g ( x1 ) = M (q ) −1 , f = − M (q ) −1 χ   −2 I n I n   In 
χ =  1 ; A =   ; D1 =  ;
(C (q, q& )q& + G (q )) , and d = M (q ) −1 J T Fc .  χ2   − I n 0n×n   0n×n 
(7)
Assumption 1: The Lipchitz condition holds for the  0n×n  %
nonlinear term as follows: D2 =   ; f = f ( x1 , xˆ 2 ) − f ( x1 , x2 )
 In 
f ( x1 , x2 + ∆x2 ) − f ( x1 , x2 ) ≤ α ∆x2 (4) Since the matrix A is negative definite, there exists a
positive definite solution P of the Lyapunov equation:
where α is a positive constant.
Assumption 2: The disturbance is limited in change rate, AT P + PA = −2 I 2 n (8)
i.e., d& ≤ δ h where δ h is a positive constant.
A Lyapunov function is selected as V = 0.5χ T P χ .
Remark 1: The main target is how to observe both joint Differentiating both sides of it, one obtains
velocities and contact force and implement them in an
admittance control framework to guarantee desired d&
admittance behavior. V& = −ωχ T χ + χ T PD1 f% − χ T PD2
ω
Remark 2: Besides the contact force, model uncertainties 1
and external disturbances always exist during the operation of ≤ −ωχ T χ + χ T χ + f% T ( PD1 )T PD1 f%
2
robot manipulators which potentially disturb the forces and (9)
1 &T
velocities estimation performance. However, this effect is + 2 d ( PD2 )T PD2d&
minor when the contact motion happens. Therefore, for 2ω
simplicity, in this paper, the effects of those difficulties are  λα2  2 1
omitted. ≤ −  ω −1 − 1  V+ λδ 2
2 2 h
 2 λ
 P1 2ω

{ } {
where λ1 = λmin ( PD1 )T PD1 , λ2 = λmin ( PD2 )T PD2 , and }
λP1 = λmax ( P) . From this, when t → ∞ , the convergence of V
is obtained as:

λ2 λP1δ h 2
V≤ (10)
 λα2 
4ω 2  ω − 1 − 1 
 2 
which leads to the estimation performance:
Fig. 1. Proposed control schematic.
δh λ2 λP1
xˆ 2 − x2 ≤
III. PROPOSED CONTROL ALGORITHM ω 3/2  1 λα2 
2λP 2 1 − − 1 
The control schematic is described in Fig. 1 which  ω 2ω 
includes the nonlinear observer, outer admittance control (11)
loop, and inner position control loop. δh λ2 λP1
dˆ − d ≤ 1/2
ω  1 λα2 
A. Proposed observer design 2λP 2  1 − − 1 
The observer is proposed in the following form:  ω 2ω 

v&1 = v2 + 2ω ( x1 − v1 ) where λP 2 = λmin ( P) . The proof of Theorem 1 is completed.

(5)
v&2 = g ( x1 )u + f ( x1 , v&1 ) + ω 2 ( x1 − v1 ) Based on Theorem 1, the contact force is estimated by
where v1 and v 2 are the internal states of the proposed Fˆc = J ( x1 ) −T M ( x1 )dˆ (12)
observer, and ω denotes the observer gain. The velocity and
B. Admittance control design
disturbance estimates are given as x̂2 = v&1 and
The outer admittance control loop is given by
dˆ = ω 2 ( x − v ) , respectively.
1 1
M d ( && pc ) + C d ( p& d − p& c ) + K d ( pd − pc ) = Fˆc
pd − && (13)
Theorem 1: The proposed nonlinear observer (5) ensures
high-accuracy estimation performances of the joint velocities where pd is the reference trajectory of the end-effector, and
and force disturbances depending on the selection of the pc is the output of the admittance controller, i.e., the
observer bandwidth.
command trajectory of the end-effector. M d , C d , and K d
Proof: Set χ1 = ( xˆ 2 − x2 ) and χ 2 = (dˆ − d ) / ω . From (3) are the mass, damper, and spring of the desired admittance
and (5), the observation error dynamics are expressed as behavior.

χ& = ω Aχ + D1 f% − D2 d& / ω (6) Based on that, the inner position control loop is designed
as follows:
where the other symbols are defined as follows:

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 01,2023 at 11:31:31 UTC from IEEE Xplore. Restrictions apply.
 α1 = p& d − k1 z1 IV. NUMERICAL SIMULATIONS
u = − M ( x1 )dˆ + C ( x1 , xˆ 2 ) + G ( x1 ) (14) In this section, a 2-DOF planar robot manipulator with
revolute joints is inherited from previous studies to verify the
−1 &
− M ( x1 ) J ( x1 ) ( J ( x1 , xˆ 2 ) xˆ 2 − α&1 + k2 zˆ2 ) effectiveness of the proposed algorithm.
where α1 is the virtual control signal following the A. Simulation setup
backstepping technique, z1 = p − pc , and zˆ2 = p&ˆ − α1 is the The mathematical model and parameters of the robot are
inherited from [9]. To simplify the validation, the reference
estimation of z2 = p& − α1 . k1 and k2 are control parameters. trajectory is set at constant, and the contact force is selected in
Theorem 2: The admittance control algorithm (14) step form. Simulation is conducted in MATLAB Simulink
guarantees that the end-effector position p tracks the environment.
command trajectory pc with arbitrarily bounded error. B. Controllers for comparison
Different admittance controllers are investigated in this
Proof: A Lyapunov function is designed as section as follows:
1 1 1) Proposed admittance controller: Outer admittance
Vc = z1T z1 + z2T z2 (15)
2 2 control parameters are chosen as M d = diag(2,2) ,
Differentiating both sides of it, it becomes C d = diag(50,50) , and K d = diag(1000,1000) . Inner
position control parameters are selected as k1 = k2 = 200 .
V&c = z1T ( p& − p& c ) + z 2T ( &&
p − α&1 ) The observer bandwidth is set as ω = 100 .
T T &
= −k1 z1 z1 + z1 z2 + z2 ( J (q, q& )q& + J (q )q&& − α&1 ) (16) 2) ESMO-based admittance controller: Both the control
T T T
configuration and the parameters are the same as the
= −k z z + z1 z2 − k2 z2 z2 + z ε Obs
1 1 1 2 proposed controller. However, the ESMO is designed to
where lumped observation errors caused by the observer are handle unknown contact force and joint velocity as follows:
given by x&ˆ1 = xˆ 2 + v
ε Obs = ( J& (q, q& )q& − J& (q, xˆ 2 ) xˆ 2 ) + J (q )(dˆ − d ) x&ˆ = g ( x )u + f ( x , xˆ ) + dˆ + 2ω v
2 1 1 2 (20)
−1
(17) &
− k2 J (q )( xˆ 2 − x2 ) + J (q) M (q) (C (q, xˆ 2 ) − C (q, q& )) dˆ = ω 2v
Since the higher the observer gain, the smaller the observer  x1 − xˆ1
error, there exists a limitation for this term which leads to η x − xˆ + ε if x1 − xˆ1 ≠ 0
v= 1 1 (21)
 1 1 0 otherwise
V&c ≤ −  k1 −  z1T z1 − ( k2 − 1) z2T z2 + ε ObsT ε Obs  n×1
 2 2 (18)
where η = 2 , ε = 0.01 , and ω = 100 .
≤ −λVc + δ

 1  1 3) ESO-based admittance controller: Both the control

where λ = min  k1 − , k2 − 1 and δ = sup ε ObsT ε Obs .
 2  2
( ) configuration and the parameters are the same as the
proposed controller. However, the ESO is designed to handle
Therefore, by some mathematical transformation, one unknown contact force and joint velocity as follows:
obtains
x&ˆ1 = xˆ 2 + 3ω ( x1 − xˆ1 )
 δ δ xˆ& = g ( x )u + f ( x , xˆ ) + dˆ + 3ω 2 ( x − xˆ )
Vc (t ) ≤  Vc (0) −  e − λt + (19) 2 1 1 2 1 1
 λ λ &
dˆ = ω 3 ( x1 − xˆ1 )
Based on this, the proof of Theorem 2 is completed.
where ω = 100 .
C. Simulation results
In Fig. 3, the contact force estimation performance of the
three observers is demonstrated. It is observed that both
observers guarantee accurate force estimation performance in
case only position of joint angles information is available, and
velocities are unknown. However, in the transient response,
the ESO shows worse force estimation performance compared
to the remaining observers since the observer order is quite
high. In contrast, the ESMO shows better transient
performance since the switching function of the error is
utilized to quickly adapt to changes in the contact force
compared to the proportional term of the ESO. In the case of
the proposed observer, since the observer order is lower than
those of the ESMO and the ESO, it not only reduces the
Fig. 2. Investigated robot model.

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 01,2023 at 11:31:31 UTC from IEEE Xplore. Restrictions apply.
 computational burden but also provides the best transient outer admittance control loop. However, in the transient time,
response in the contact force estimation. the proposed controller and the ESMO dominates the ESO in
term of fast and accurate response when the contact motion
happens. Besides, it is interesting that the proposed observer
and the ESMO provide almost equivalent tracking
performance, which proves the superiority of the proposed
FcX (N)

observer with fewer parameters and complexity in the

observer design. In Fig. 5, the control signals of both
controllers are given. One observes that both controllers
generate smooth control signals while the magnitude of the
proposed control signal in the transient is lower than the
remaining thanks to the high-accuracy contact force and
velocity estimation of the proposed observer.
FcY (N)

V. CONCLUSION
In this paper, the problem of force sensorless admittance
control for robot manipulators is investigated by introducing a
novel observer to estimate both contact force and joint
velocities. The proposed observer guarantees not only
accurate estimation performance in the steady state but also a
Fig. 3. Interaction force estimation.
fast response in the transient time. Based on this, an
admittance controller is constructed, which guarantees the
0.47
Commanded ESO ESMO Proposed
desired admittance behavior of the control system.
0.46 0.46
REFERENCES
p X (m)

0.45 0.46
0.45
0.44 0.45 [1] R.-D. Xi, X. Xiao, T.-N. Ma, and Z.-X. Yang, "Adaptive Sliding Mode
0.44 4 4.05 4.1 4.15 4.2
Disturbance Observer Based Robust Control for Robot Manipulators
0.44
16 16.1 16.2
Towards Assembly Assistance," IEEE Robotics and Automation
0 5 10 15 20
Letters, vol. 7, no. 3, pp. 6139-6146, 2022, doi:
Time (seconds) 10.1109/lra.2022.3164448.
[2] Y. Sun, J. Falco, M. A. Roa, and B. Calli, "Research Challenges and
0.47 Commanded ESO ESMO Proposed Progress in Robotic Grasping and Manipulation Competitions," IEEE
Robotics and Automation Letters, vol. 7, no. 2, pp. 874-881, 2022, doi:
0.46
10.1109/lra.2021.3129134.
p Y (m)

0.46
0.46
0.45 [3] B. Wang, S. J. Hu, L. Sun, and T. Freiheit, "Intelligent welding system
0.45 0.45 technologies: State-of-the-art review and perspectives," Journal of
0.44
0.44
16 16.1 16.2 Manufacturing Systems, vol. 56, pp. 373-391, 2020/07/01/ 2020, doi:
0.44
4 4.05 4.1 4.15 4.2 https://doi.org/10.1016/j.jmsy.2020.06.020.
0 5 10 15 20 [4] H. V. Dao and K. K. Ahn, "Extended Sliding Mode Observer-Based
Time (seconds) Admittance Control for Hydraulic Robots," IEEE Robotics and
Automation Letters, vol. 7, no. 2, pp. 3992-3999, 2022, doi:
Fig. 4. End-effector position.
10.1109/lra.2022.3147244.
[5] N. Hogan, "Impedance Control: An Approach to Manipulation: Part
30 I—Theory," Journal of Dynamic Systems, Measurement, and Control,
20
ESO ESMO Proposed vol. 107, no. 1, pp. 1-7, 1985, doi: 10.1115/1.3140702.
0 [6] S. H. Chien, J. H. Wang, and M. Y. Cheng, "Performance Comparisons
10
-10
of Different Observer-Based Force-Sensorless Approaches for
0 10 Impedance Control of Collaborative Robot Manipulators," in 2020
-20
-10 4 4.05 4.1 4.15 0
IEEE Conference on Industrial Cyberphysical Systems (ICPS), 10-12
-10
-20 16 16.05 16.1 16.15
June 2020 2020, vol. 1, pp. 326-331, doi:
10.1109/ICPS48405.2020.9274790.
0 5 10 15 20
Time (seconds)
[7] S. Lee, M. Kim, and J. Song, "Sensorless collision detection for safe
human-robot collaboration," in 2015 IEEE/RSJ International
10
ESO ESMO Proposed
Conference on Intelligent Robots and Systems (IROS), 28 Sept.-2 Oct.
0 2015 2015, pp. 2392-2397, doi: 10.1109/IROS.2015.7353701.
-5 [8] J. Yang and C. Peng, "Adaptive neural impedance control with
0
-10 5 extended state observer for human–robot interactions by output
4 4.05 4.1 4.15
0 feedback through tracking differentiator," Proceedings of the
-5
-10 Institution of Mechanical Engineers, Part I: Journal of Systems and
16 16.05 16.1 16.15 Control Engineering, vol. 234, no. 7, pp. 820-833, 2020, doi:
0 5 10 15 20 10.1177/0959651819898936.
Time (seconds) [9] S. Zhang, P. Yang, L. Kong, W. Chen, Q. Fu, and K. Peng, "Neural
Networks-Based Fault Tolerant Control of a Robot via Fast Terminal
Fig. 5. Control signal. Sliding Mode," IEEE Transactions on Systems, Man, and Cybernetics:
Systems, vol. 51, no. 7, pp. 4091-4101, 2021, doi:
The admittance behaviors of the three controllers are 10.1109/tsmc.2019.2933050.
expressed in Fig. 4. Similar to the contact force estimation, in
the steady state, both controllers show high-accuracy tracking
performance with the commanded trajectory from the ideal

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 01,2023 at 11:31:31 UTC from IEEE Xplore. Restrictions apply.