2016 IEEE 55th Conference on Decision and Control (CDC)

ARIA Resort & Casino

December 12-14, 2016, Las Vegas, USA

MPC-based Admittance Control for Robotic Manipulators

Arne Wahrburg, Kim Listmann

Abstract— Robotic applications involving environmental in- admittance control [7], [8] or by introducing oscillation
teraction require control structures that go beyond traditional cancellation to the closed-loop [9]. The fundamental trade-
position control. One approach to handle interaction forces is off in admittance control is the interplay of velocity of the
admittance control. Therein, position or speed reference values
are modified based on interaction forces. To this end, the operation (including at and under contact) and the strength
underlying position control scheme, widely used in industrial of the generated contact forces. Ideally, high velocity (=
manipulators, is augmented with an additional admittance productivity) is paired with low contact forces to protect
control loop. It is well known that such approaches raise the robot and the environment from damage. To this end,
stability issues in the interaction with stiff environments (termed adaptive admittance control laws limiting contact forces
contact stability). As a consequence, in traditional admittance
control, the speed of operation needs to be reduced in order were developed by [10], while [11] prevented bouncing by
to prevent excessive contact force magnitudes and to ensure applying sliding mode control.
stability. In this paper, we propose an MPC-based admittance We will propose a different control approach here. Given
control scheme to circumvent both and discuss its properties the desire to constrain contact forces while staying in a de-
and difficulties for practical use. fined contact with the environment, a natural way to describe
I. I NTRODUCTION this is by using a model-predictive control (MPC) framework.
Hence, the main contribution of this article is to present and
Since its foundations in [1] impedance and admittance discuss an MPC admittance control scheme. While MPC
control reached industrial maturity. The main driver for that has been used for robot motion control in [12], [13] this
is actually the trend towards a more interactive usage of is the first implementation related to an interaction control
robots even in industrial environments [2]. While formerly scheme for industrial robots. It is applicable to force-torque
robots where intended to be used as highly effective means of based interactions of the robot TCP with the environment, for
production, separated from environmental interaction behind all (selected) environmental (Cartesian) degrees of freedom
fences, the age of collaborative robots shows that the new (DoF). In this paper, we will, for the sake of clarity and
form of use is much more interactive - with humans as much brevity, rely on some simplifying assumptions that help us
as with the surrounding environment. focus on the development of the general method. We are
This must be reflected in the control schemes of the robots well aware of the fact that such assumptions pose additional
and necessitates the use of force control techniques over challenges to the method that need to be investigated in the
classical motion control. If the particular interest is in smooth future.
environmental interaction, impedance or admittance control The remainder of the paper is organized as follows: In the
are key [3]. While the former creates a force correction term next section the problem solved in this article is properly
based on velocity deviations both in Cartesian space, the introduced together with some notation. Then we show the
opposite holds for the latter. This directly determines the general approach of MPC admittance control in Section III
main terms of use for both methods [3]: (i) Impedance con- and provide first simulation results for a 2 DoF planar robot
trol is mainly intended for stiff behavior and inertia shaping in Section IV. We address real-world effects in Section V
with the potential to demand high gains in the control loop. and point out various directions for future research based on
Hence, the fundamental problems are noise amplification and the concept proposed in this paper. Finally, we conclude the
instability in closed loop operation and friction dependent article in Section VI.
accuracy. (ii) Admittance control, on the contrary, tends
to instability in stiff contact situations [4, Section 1.3.2], II. P RELIMINARIES
but combines well with cascaded, existing industrial motion We consider rigid-link, stiff-joint robotic manipulators
control schemes of the robots since it is relying on the described by the dynamics
feed-forward decoupling [5, Section 7.2]. Hence, admittance
control features some industrial applications mentioned in M (q)q̈ + C(q, q̇)q̇ + G(q) + τfric + τext = τmot , (1)
[3] and references therein. with q ∈ RNdof where Ndof is the number of degrees of
Despite first industrial use, the stability problems at stiff freedom of the manipulator. In (1), M q̈ describes inertial
interaction situations, first reported in [6], remain. Later torques, C q̇ captures centrifugal and Coriolis effects, and G
people tried to circumvent those by formulating adaptive are joint torques resulting from gravity. Furthermore, τfric are
friction torques in the joints, τext are joint torques induced
The authors are with the ABB AG, ABB Corporate Research
Center Germany, Wallstadter Str. 59, 68526 Ladenburg, GERMANY by external loads (e.g. environmental interaction), and τmot
{arne.wahrburg;kim.listmann}@de.abb.com. are the motor torques driving the manipulator.

978-1-5090-1837-6/16/$31.00 ©2016 IEEE 7548

Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.
 f In case the TCP hits a stiff surface with high speed, the
values in D have to be very small in order to generate
ṗ0 ṗref q̇ref vel. τmot q, q̇ large magnitudes in ∆ṗ to slow down the manipulator and
IK Robot
+ control prevent excessive contact forces. However, small values in D
may cause a weakly damped behavior which often results in
imot loosing contact to the environment (also denoted as contact
F/T est. instability). Several extensions have been developed to tackle
the aforementioned problem of (3), including adaptive admit-
∆ṗ adm. fˆ tance controllers [7], [8], [10], sliding-mode approaches [11]
control and other non-linear extensions [9].
However, to the best of the authors’ knowledge, none of
the existing approaches fully accomplish the two fundamen-
Fig. 1. Overview of Cartesian admittance control – The manipulator is
controlled by a velocity controller at joint level. An outer control loop tal requirements for any admittance control:
generates Cartesian speed offsets ∆ṗ based on estimates fˆ of external 1) Contact forces and torques must not exceed specified
wrench. Together with the originally commanded Cartesian speed reference
ṗ0 they are mapped to joint space by means of inverse velocity kinematics
thresholds to prevent damage to both, the environment
resulting in the joint speed reference values q̇ref . The closed-loop velocity and the manipulator.
dynamics are marked by the dashed box. 2) Absolute values of contact forces and torques must not
become zero to ensure steady contact with the environ-
ment and prevent contact instability or bouncing.
Throughout the paper, we assume that environmental con-
Based on the formulation of these two requirements and a
tact with the manipulator may only occur at the TCP. While
simplified manipulator and environmental model, a model-
this might seem restrictive, it is a reasonable assumption
predictive approach to admittance control seems natural.
for robotic applications such as assembly or grinding. The
With MPC being well-known for its excellent capabilities
Cartesian contact forces and torques exchanged between the
of explicitly handling both input and state constraints, we
TCP and the environment are summarized into
T propose an MPC-based admittance control law.
f = fx , fy , fz , τx , τy , τz ∈ RNcart ,

(2)
III. MPC- BASED A DMITTANCE C ONTROL
which is typically referred to as wrench. Under the aforemen- Numerous approaches to control the position of a ma-
tioned assumption of contact only occurring at the TCP, the nipulator described by (1) have been reported in literature,
external load torques τext are linked to the wrench via τext = including centralized PD with gravity compensation [16],
J T (q)f , which can be deduced by the principle of virtual computed torque [17], and adaptive control [18], only to
work. Therein, J(q) is the manipulator Jacobian. In general, name a few. In industrial manipulators, independent joint
Ncart = 6 holds as in (2). In some situations it is sufficient to control combined with decoupling feed-forward [19] is the
consider a reduced dimension of wrench. This includes pure most common scheme due its high performance and robust-
point contacts with the environment where no contact torques ness in position control [5, Section 7.2].
occur (Ncart = 3), general planar manipulators (Ncart = 3) Assuming that such a control scheme is well-tuned, the
and a combination of both (Ncart = 2). closed-loop dynamics of each joint can be described by a
Traditionally, industrial manipulators are position or speed very simple model. Given the joint speed reference qref,i
controlled at joint level. In a Cartesian or operational space of joint i ∈ {1, . . . , Ndof }, the closed-loop dynamics of the
admittance control approach, such a joint level control speed loop can be approximated by
scheme is augmented with an additional loop modifying
Cartesian speed reference values. Let p, ṗ denote the Carte- 1
q̇i = q̇ref,i , ∀i = 1, . . . , Ndof , (4)
sian position vector and its velocity, then the above modifi- TV,i s + 1
cation results from with time constants TV,i . In contrast to the common assump-
∆ṗ = −D −1
f (3) tion q̇i = q̇ref,i [5, Section 7.2], the model (4) takes into
account the limited bandwidth of velocity control and has
with a diagonal damping matrix D [5, Section 7.2]. The been verified by experimental data [20].
modified Cartesian speed reference is then mapped to joint Based on (4), a system model for designing an MPC-based
a speed reference employing the inverse velocity kinemat- admittance control will be deduced in the following.
ics of the manipulator (i.e. an inverse of the manipulator
Jacobian). The approach described above is depicted in A. A System Model for MPC-based Admittance Control
Fig. 1, where an additional subsystem for estimating external As depicted in Figure 1, the admittance control scheme
wrench based on motor currents is included [14]. In case modifies Cartesian speed reference values according to ṗref =
additional force/torque sensing is available, the wrench f ṗ0 + ∆ṗ, where ṗ0 is the original Cartesian speed reference
can directly be fed back to the admittance controller. and ∆ṗ is the corresponding offset. Hence, the output of the
While the admittance scheme (3) is intuitive, it suffers admittance controller, ∆ṗ, is interpreted as the control signal
from a major drawback as pointed out e.g. in [3], [15]: for the velocity controlled robot.

7549
Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.
 As pointed out in (4), the dynamics of the velocity- describing the environment as a spring. Denoting the forward
controlled manipulator can be described by a set of decou- kinematics of the manipulator by Ψ (q), the insertion depth
pled first-order lag elements. Thus, the velocity dynamics of with respect to the initial contact point pcontact is given
the entire manipulator in Laplacian domain are given by by Ψ (q) − Ψ (qcontact ). With Ke describing environmental

1 1
 stiffness, contact wrench is given by
q̇ = diag ,..., q̇ref . (5)
TV,1 s + 1 TV,Ndof s + 1 f = Ke (Ψ(q) − Ψ(qcontact )) . (11)
Apart from the manipulator model, information about contact
Formulating (6) in time domain, introducing
forces and torques is obviously a key element for admittance
 
control. Those quantities may be obtained from a wrist 1 1
mounted force/torque sensor, for example. However, such ΛTest = diag ,..., , (12)
Test,1 Test,Ncart
additional sensing substantially increases the cost of the
robotic system and is thus not available in many cases. An and substituting (11), the dynamics of fˆ can be written as
alternative to measuring wrench is estimating its compo-
nents from motor currents and a manipulator model. While ẋ3 = −ΛTest x3 + ΛTest Ke (Ψ(q) − Ψ(qcontact )) . (13)
different approaches to contact force estimation exist (see Remark 2: Note that throughout the paper we assume
e.g. [14], [21], [22], [23]), the resulting components of pcontact is known to apply our scheme. In practice this
the wrench estimate can be well approximated by low-pass is tricky, of course, and involves accurate environmental
filtered versions of the actual contact forces and torque. perception and scene interpretation. However, we focus on
Therefore, we rely on developing the basic idea of MPC admittance control here
 
1 1 and interpret such practical obstructions as challenges for the
fˆ = diag ,..., f (6)
Test,1 s + 1 Test,Ncart s + 1 future of the method.
The equations ẋ1 = ẋ2 , (10), and (13) define the system
for the purpose of designing a system model that enables
model that is employed for designing the MPC-based ad-
MPC-based admittance control.
mittance controller. Based on a proper discretization, e.g. by
Combining joint positions, joint speeds, and wrench we
simple first-order Euler with sampling time Ts , the system
define the state vector as
    dynamics can be summarized into
x1 q
x = x2  =  q̇  ∈ R2Ndof +Ncart . (7) xk+1 = Axk + B k ∆ṗk + E k ṗ0 + γ k , (14a)
x3 fˆ
 
IN Ts INdof 0
The dynamics of x1 are trivially given by ẋ1 = x2 . In A =  0 INdof − Ts ΛTV 0  , (14b)
order to express the dynamics of x2 (i.e. joint accelerations) 0 0 INcart − Ts ΛTest
 
in terms of the commanded Cartesian speed offset ∆ṗ, we 0
employ the inverse velocity kinematics of the manipulators. B k = Ts ΛTV J −1 (xk1 ) , (14c)
Assuming non-singular configurations, this is described by 0

q̇ = J −1 (q)ṗ (8) Ek = Bk , (14d)

 
0
for traditional 6 DoF manipulators. γk =  0
. (14e)

Remark 1: We emphasize that the approach presented in Ts ΛTest Ke Ψ(xk1 ) − Ψ(qcontact )
this paper can readily be extended to redundant manipulators.
The proposed MPC-based admittance control scheme oper- Note that corresponding to Fig.1 ∆ṗ is considered a control
ates in Cartesian space and generates speed offsets ∆ṗ. For input in our case. While it is conceivable to design an MPC-
redundant manipulators, the mapping (8) is to be replaced by based control law directly for (14), next we introduce further
an arbitrary redundancy resolution scheme mapping ∆ṗ onto model simplifications based on practical observations that
∆q̇. For a recent overview we refer to [24] but constrain our allow applying linear MPC schemes.
considerations to non-redundant manipulators for the sake of
brevity and clarity. B. Model Simplifications for Application of Linear MPC
Transforming (5) into time domain, introducing The system model (14) contains the inverse Jacobian J −1

1 1
 and the forward kinematics Ψ, both being functions of the
ΛTV = diag ,..., , (9) joint configuration x1
= q. To obtain a linear system model,
TV,1 TV,Ndof
we assume J −1 xk1 to be constant over the prediction
and substituting ṗref = ṗ0 + ∆ṗ into (8), we obtain horizon of the MPC (but updated every time step, of course).
ẋ2 = −ΛTV x2 + ΛTV J −1 ṗ0 + ΛTV J −1 ∆ṗ. (10) This is motivated by the fact that the configuration will
only be subject to small changes over the prediction horizon
In order to model the relation between wrench and TCP as long as the manipulator is kept away from singular
pose, we rely on the simple yet well-established model of configurations. The forward kinematics in turn are linearized

7550
Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.
about an operating point x1 , which is the joint configuration In order to only activate certain Cartesian directions for
at the beginning of the prediction horizon. This results in admittance control, the dimension of the input u = ∆p ∈
Rκadm can be reduced to only describe admittance con-
Ψ(xk1 ) ≈ Ψ(x1 ) + J(x1 ) xk1 − x1 .


(15)
trolled directions. Consequently, only forces in the respective
Based on the aforementioned assumptions, the relation directions have to be estimated and the Jacobian in the
between wrench and joint configuration is approximated as model (18) is reduced to the rows corresponding to the
admittance controlled directions. Thereby, purely position-
flin = Ke Ψ(x1 )− Ψ(qcontact )− J(x1 )x1 + J(x1 )xk1 .


controlled Cartesian directions remain unaffected by the
(16) admittance controller.
Thus, the system model (14) can be simplified to the affine
Due to the constraint (18e), the control input resulting
system
from (18) cannot be used when the TCP moving through free
xk+1 = Axk + B∆ṗk + E ṗ0 + γ, space. We propose using thresholds on the force estimates

INdof Ts INdof 0
 fˆ to trigger activation of the admittance controller. If an
A= 0 IN −Ts ΛTV 0 , estimate in a certain direction is below a specified threshold
Ts ΛTest Ke J(x1 ) 0 INcart −Ts ΛTest β, the corresponding direction is removed from the set of
  admittance controlled directions. For free motion, the set of
0
admittance controlled directions is thus empty. Note that this
B = E = Ts ΛTV J −1 (x1 ) , (17)
removal impacts the constraints (18d), (18e), since they do
0
  not need to hold for the removed Cartesian directions.
0 On the one hand, the threshold β should be chosen large
γ = 0 . enough to ensure the admittance controller does not react to
Ts ΛTest Ke (Ψ(x1 )−Ψ(qcontact )−J(x1 )x1 ) noise and/or uncertainties in the wrench estimates. On the
We emphasize that in contrast to (14), the input matrices B other hand, a too large threshold β delays the activation of
and E are constant as well as the affine term γ. the admittance controller. Due to the limited bandwidth of
Based on (17) we formulate the MPC problem the speed control loop (4) and the fact that forces may rise
H−1
quickly in stiff contact situations, the MPC problem (18)
min
X
xi − x0

T T
Q xi − x0 + ∆ṗi R∆ṗi


(18a) may be infeasible. From a practical perspective, a feasibility
∆ṗ
i=1
check for (18) has to be performed first if the threshold β is

T exceeded and admittance control is activated. This takes into
+ xH − x0 Q xH − x0 ,


(18b) account estimated contact forces, manipulator joint speeds,
s.t. (17), (18c) given contact force constraints, the time constants describing
|flin,j | ≤ |fmax,i | , ∀j = 1, . . . , Ncart , (18d) the independent joint control, and environmental stiffness.
In case the MPC problem is infeasible, an error triggers the
|flin,j | > 0, ∀j = 1, . . . , Ncart , (18e)
manipulator to stop its motion.
where the prediction horizon H is used. The set point Remark 4: Please note that for the implementation we
x0 = [q0 , q̇0 , 0]T contains the originally commanded joint rely on state-of-the-art suites (e.g. FORCES [25], ACADO
configuration and speeds and is constant over the prediction [26]) for obtaining a solution to (18). These automatically
horizon. condense dynamic constraint into a single equality constraint
Remark 3: Assigning a wrench of f0 = 0 is no contra- and provide feasbility checks of the given problem.
diction to establish steady contact forces in environmental
interaction. Thanks to the assumption of spring-like behavior IV. S IMULATION E XAMPLE
of the environment, any penetration results in a Cartesian
wrench. To highlight the benefits of the proposed MPC-based
The constraints (18d) and (18e) reflect the two fun- admittance control, we apply it to the planar 2 DoF ma-
damental requirements for stable admittance control: The nipulator with Ndof = Ncart = 2 sketched in Fig. 2. Starting
magnitudes of contact forces have to be limited and must from an initial configuration, joint speed reference values q̇0
not become zero (avoid damage and prevent bouncing). are commanded to move the TCP in negative y-direction.
During this motion, the TCP hits an inclined surface. The
C. Admittance Control in Specific Cartesian Directions environment is described by the stiffness matrix expressed
The MPC problem (18) applies for situations in which all in the coordinates of the inclined surface as
Cartesian directions are admittance controlled and the TCP  
0 0
exchanges contact forces and torques with the environment Ke,plane = . (19)
0 kenv
in all Cartesian directions. In a lot of robotic applications,
the former is not the case (including e.g. free robot motion). Denote the surface tilt angle by α, then the stiffness matrix
As an example, consider tracking a trajectory in the xy-plane in base-frame coordinates is given by
on an uneven surface. Then, admittance is only intended in
T
z-direction. Ke = Tplane Ke,plane Tplane (20)

7551
Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.
 0.8 This is also visualized in Figure 5. While the standard
scheme with low damping bounces along the surface instead
0.6 of sliding, the standard scheme with high damping penetrates
y in m

the surface depply resulting in high forces. On the contrary,

0.4
the MPC-based admittance controller maintains a small but
0.2 non-zero penetration depth and smoothly slides along the
surface.
0 All control laws run at a sampling rate of Ts = 4 ms and
0 0.5 1 1.5
we emphasize that the commanded speed of operation is the
x in m same for all cases.
Fig. 2. Application of MPC-based admittance control to a 2D planar
manipulator – The TCP is commanded to move in negative y-direction,
V. D ISCUSSION OF R ESULTS AND F UTURE W ORK
with admittance control enabled in both x- and y-direction. Upon hitting Standard admittance control is known to suffer from
the inclined surface, the TCP slides down along the surface.
problems in stiff environmental contact [6]. To ensure stable
contact, the dampings in (3) have to be large, often resulting
with in high contact forces. Due to this limit, speed of operation
typically has to be reduced. The proposed MPC-based ad-
 
cos(α) − sin(α)
Tplane = . (21) mittance control scheme solves this problem and comes with
sin(α) cos(α)
several benefits. Explicit limits on tolerated contact forces
In the example, we use α = 25◦ and ke = 105 N m−1 . can be specified and non-zero contact force magnitude are
Furthermore, the joint level speed controllers are described guaranteed. Thus, stable contact is achieved with bounded
by TV,1 = 15ms and TV,2 = 10ms and force estimation is force magnitudes. Modern industrial manipulators achieve
assumed to have the time constant Test,i = 20ms, i = 1, 2. very high performance and accuracy in joint position and
The objective is to achieve stable environmental contact speed control. Even though manipulator dynamics are non-
without exceeding a contact force magnitude of |fmax,y | = linear and coupled, a remarkably simple linear model based
200N. on (5) is sufficient to achieve very promising results using
First, two standard admittance controllers according to MPC-based admittance.
(3) are employed to handle the contact situation. These We also showed that the MPC-based control can explicitly
controllers are parameterized by take into account the additional dynamics introduced by
estimating external wrench. Due to the predictive nature of
Dlow = diag (100, 100) , Dhigh = diag (1000, 2000) . (22)
the proposed scheme, extensions towards force estimation or
The damping values have been manually tuned to meet one measurement with significant time delays is a topic to be
of the two requirements in each case. As the evaluation of tackled in future research. Identified in [6] as a reason for
the contact force in y-direction over time in Figure 3 shows, instability in standard admittance control schemes, taking
Dlow roughly satisfies the constraint on the maximum contact
force.
0
However, the contact force cyclically becomes zero as
fy in N

the TCP bounces along the inclined surface. Using the high -200
damping values Dhigh , contact is maintained since |fy | > 0 -400
is ensured. However, the tolerated magnitude of the contact
-600
force is exceeded by more than 200%.
0.8 1 1.2 1.4 1.6 1.8 2
In comparison to the two standard admittance control
time in s
schemes, the proposed MPC-based admittance control is sim-
ulated. The prediction horizon is H = 5 and the weighting 1.5
∆ẏ in m s−1

matrices are chosen as 1

Q = diag (1000, 1000, 1000, 1000, 0.001, 0.001), (23a) 0.5
R = diag (10, 10) . (23b) 0
0.8 1 1.2 1.4 1.6 1.8 2
A force threshold of β = 5 N is chosen for activation of the
time in s
admittance control scheme. As visualized in Fig. 3 and the
close-up Fig. 4, the MPC-scheme meets both requirements. Fig. 3. Contact force and Cartesian speed offset in y-direction for different
The proposed scheme produces large Cartesian speeds offsets admittance control schemes – The dashed line shows the result using small
in the initial contact phase that rapidly slow down the damping values in a standard admittance control scheme. The maximum
tolerated contact force magnitude is not exceeded but stable contact cannot
manipulator motion in y-direction and prevent contact forces be maintained. The dash-dotted line shows the result with high dampings.
from becoming too high. Furthermore, stable steady contact While stable contact is maintained, excessive contact forces are observed.
is maintained. The MPC-based admittance control (solid line) meets both requirements.

7552
Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.
 0 relies on a model of the environment. In real-world scenarios,
fy in N
−100 the behavior of the environment will always be subject
to uncertainties. To this end, it is conceivable to integrate
−200
online identification schemes for local approximations of
the environmental model [27], [28], [29], [30], [31], [32]
0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96
into the proposed admittance control scheme. Related to the
time in s
interaction with the environment, the joint configuration of
1 initial contact, qcontact is in general also not known a-priori.
∆ẏ in m s−1

An ad-hoc approach towards this problem is to define qcontact

0.5 as the configuration at which the activation threshold β is
exceeded for the first time. In case of estimated contact
0 forces and torques, this would however introduce an offset
0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 due to the lag introduced by wrench estimation. Hence, more
time in s sophisticated methods might be needed (see e.g. [27]).
Recently, considerable efforts have been devoted to adap-
Fig. 4. Details of initial contact phase – The MPC-based admittance
controller (solid line) meets the contact force constraint |fy | ≤ 200 N and
tive impedance control for intuitive physical human-robot
maintains stable contact (i.e. |fy | > 0 N). This is achieved by the fast initial interaction (pHRI). In pHRI, bounding contact forces is
response as visible from ∆ẏ at t = 0.82 s. The standard admittance scheme obviously of paramount importance for safety reasons. Fur-
with low dampings (dashed) looses contact whereas high dampings result
in exceeding the specified maximum contact force magnitude.
thermore, the virtual variation of stiffness and damping
described in [9], [33], [34], [35] is intended to achieve
intuitive interaction. Another direction for future research is
to quantify measures for intuition (e.g. energy transferred
0.68 between human and robot during interaction) into the cost
y in m

function of an MPC problem and to investigate the applica-

0.66 bility of schemes similar to the one proposed in this paper
for pHRI.
0.64 Another aspect worth mentioning is the parameterization
of the MPC-control law, i.e. tuning the weighting matrices Q
1.32 1.34 1.36 1.38 1.4 and R and the prediction horizon H. Compared to specifying
x in m the contact force constraints, this task is less intuitive and
from a practical perspective, an automatic parameterization
Fig. 5. Close-up of TCP trajectories in initial contact phase – The standard would be desirable.
admittance control law with low damping (dashed line) bounces of the
surface. With high damping (dash-dotted), the standard scheme penetrates VI. C ONCLUSION
the surface to a large amount. The proposed MPC-based controller ensure
stable contact with small surface penetration. We presented a novel approach to admittance control for
robotic manipulators. The key benefits of the MPC-based
admittance controller are that contact forces and torques can
time delays in force determination into account systemat- be explicitly limited and are guaranteed to be non-zero in
ically might turn out to be one of the major advantages magnitude ensuring stable contact without bouncing even in
of the proposed new MPC approach to admittance control. stiff environmental contact. Furthermore, it was shown that a
Nonetheless, it is possible to apply the scheme to systems remarkably simple linear model is sufficient to describe the
with ideal force measurement, i.e. fˆ = f , of course. In dynamics of a velocity controlled manipulator and apply the
that case, the dynamics of wrench estimation can be omitted newly developed concept.
and the system model employed for MPC design is further The new concept is considered a starting point for numer-
simplified. ous further investigations, ranging from real-time implemen-
An immediate next step is experimental validation. Since tation and experimental verification to online identification
the overall dimension of the state-space is 2Ndof + Ncart of environmental models and MPC-based admittance control
and thanks to the linear nature of the system model (17) laws for physical human-robot collaboration.
and the short prediction horizon, a real-time implementation
e.g. using [25], [26] seems feasible. Furthermore, it is an R EFERENCES
open question if more sophisticated models (e.g. (14) without [1] N. Hogan, “Impedance control: An approach to manipulation (part i-
linearizations) are worth the additional computational effort iii),” ASME Journal of Dynamic Systems, Measurement, and Control,
vol. 107, pp. 1–24, 1985.
they impose. [2] R. Drath, B. Matthias, A. Horch, M. W. Krüger, and K. D. Listmann,
From a practical perspective, at least two main challenges “The Internet of Things, Services and People,” atp edition, vol. 58,
have been identified: First, the MPC-based admittance con- pp. 36–43, Mar. 2016.
[3] C. Ott, R. Mukherjee, and Y. Nakamura, “A hybrid system framework
trol scheme is based on a model of the manipulator, which for unified impedance and admittance control,” Journal of Intelligent
is well known, but secondly the system description (17) also and Robotic Systems, vol. 78, no. 3, pp. 359–375, 2015.

7553
Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.
 [4] C. Ott, Cartesian impedance control of redundant and flexible-joint programming,” Mathematical Programming Computation, vol. 6,
robots. Springer, 2008. no. 4, pp. 327–363, 2014.
[5] B. Siciliano and O. Khatib, eds., Springer Handbook of Robotics. [27] A. Stolt, M. Linderoth, A. Robertsson, and R. Johansson, “Adaptation
Springer, 2008. of force control parameters in robotic assembly,” in Proc. of IFAC
[6] D. Lawrence, “Impedance control stability properties in common im- International Symposium on Robot Control, pp. 561–566, 2012.
plementations,” in Proc. of IEEE International Conference on Robotics [28] D. Erickson, M. Weber, and I. Sharf, “Contact stiffness and damping
and Automation, pp. 1185–1190, 1988. estimation for robotic systems,” International Journal of Robotics
[7] J. Roy and L. Whitcomb, “Adaptive force control of position/velocity Research, vol. 22, no. 1, pp. 41–57, 2003.
controlled robots: Theory and experiments,” IEEE Transactions on [29] D. Verscheure, I. Sharf, H. Bruyninckx, J. Swevers, and J. De Schut-
Robotics and Automation, vol. 18, no. 2, pp. 121–137, 2002. ter, “Identification of contact parameters from stiff multi-point con-
[8] H. Wang, F. Patota, G. Buondonno, M. Haendl, A. De Luca, and tact robotic operations,” International Journal of Robotics Research,
K. Kosuge, “Stability and variable admittance control in the physical vol. 29, no. 4, pp. 367–385, 2010.
interaction with a mobile robot,” International Journal of Advanced [30] A. Haddadi and K. Hashtrudi-Zaad, “Online contact impedance iden-
Robotics, vol. 12, no. 173, pp. 1–14, 2015. tification for robotic systems,” in Proc. of IEEE/RSJ International
[9] P. Labrecque and C. Gosselin, “Robotic force amplification with free Conference on Intelligent Robots and Systems, pp. 974–980, 2008.
space motion capability,” in Proc. of IEEE International Conference [31] L. Roveda, F. Vicentini, and L. M. Tosatti, “Deformation-tracking
on Robotics and Automation, pp. 134–140, 2014. impedance control in interaction with uncertain environments,” in
Proc. of IEEE/RSJ International Conference on Intelligent Robots and
[10] C.-H. Liang, S. Bhasin, K. Dupree, and W. Dixon, “A force limiting
Systems, pp. 1992–1997, 2013.
adaptive controller for a robotic system undergoing a noncontact-to-
[32] L. Roveda, Model Based Compliance Shaping Control of Light-Weight
contact transition,” IEEE Transactions on Control Systems Technology,
Manipulator in Hard-Contact Industrial Applications. PhD thesis,
vol. 17, no. 6, pp. 1330–1341, 2009.
Politecnico di Milano, 2015.
[11] R. Kikuuwe, “A sliding-mode-like position controller for admittance
[33] R. Ikeura and H. Inooka, “Variable impedance control of a robot for
control with bounded actuator force,” IEEE/ASME Transactions on
cooperation with a human,” in Proc. of IEEE International Conference
Mechatronics, vol. 19, no. 5, pp. 1489–1500, 2014.
on Robotics and Automation, pp. 3097–3102, 1995.
[12] D. Verscheure, B. Demeulenaere, J. Swevers, J. D. Schutter, and [34] F. Ficuciello, L. Villani, and B. Siciliano, “Variable impedance control
M. Diehl, “Time-optimal path tracking for robots: A convex opti- of redundant manipulators for intuitive human-robot physical interac-
mization approach,” IEEE Transactions on Automatic Control, vol. 54, tion,” IEEE Transactions on Robotics, vol. 31, no. 4, pp. 850–863,
pp. 2318–2327, Oct. 2009. 2015.
[13] P. Milosavljevic, T. Faulwasser, A. Marchetti, and D. Bonvin, “Time- [35] I. Ranatunga, S. Cremer, D. Popa, and F. Lewis, “Intent aware
optimal path-following operation in the presence of uncertainty,” in adaptive admittance control for physical human-robot interaction,” in
Proceesdings of the European Control Conference 2016, p. submitted, Proc. of IEEE International Conference on Robotics and Automation,
2016. pp. 5635–5640, 2015.
[14] A. Wahrburg, E. Morara, G. Cesari, B. Matthias, and H. Ding,
“Cartesian contact force estimation for robotic manipulators using
Kalman filters and the generalized momentum,” in Proc. of IEEE
International Conference on Automation Science and Engineering,
pp. 1230–1235, 2015.
[15] A. Albu-Schäffer and G. Hirzinger, “Cartesian impedance control
techniques for torque controlled light-weight robots,” in Proc. of IEEE
International Conference on Robotics and Automation, pp. 657–663,
2002.
[16] R. Kelly, “PD control with desired gravity compensation of robot
manipulators: A review,” International Journal of Robotics Research,
vol. 16, no. 5, pp. 660–672, 1997.
[17] F. Lewis, C. Abdallah, and D. Dawson, Control of Robot Manipulators.
Macmillan, 1993.
[18] S. Dubowsky and D. DesForges, “The application of model-reference
adaptive control to robotic manipulators,” ASME Journal of Dynamic
Systems, Measurement, and Control, vol. 37, no. 1, pp. 106–110, 1992.
[19] M. Spong and M. Vidyasagar, Robot dynamics and control. John
Wiley & Sons, 2008.
[20] F. Hans, “Compliant control for robust assembly with redundant
manipulators in partially unstructured environments,” Master’s thesis,
TU Munich, 2015.
[21] A. Stolt, M. Linderoth, A. Robertsson, and R. Johansson, “Force
controlled robotic assembly without a force sensor,” in Proc. of IEEE
International Conference on Robotics and Automation, pp. 1538–1543,
2012.
[22] H. Cho, M. Kim, H. Lim, and D. Kim, “Cartesian sensor-less force
control for industrial robots,” in Proc. of IEEE/RSJ International
Conference on Intelligent Robots and Systems, pp. 4497–4502, 2014.
[23] A. Colome, D. Pardo, G. Alenya, and C. Torras, “External force
estimation during compliant robot manipulation,” in Proc. of IEEE
International Conference on Robotics and Automation, pp. 3535–3540,
2013.
[24] A. Colome and C. Torras, “Closed-loop inverse kinematics for re-
dundant robots: Comparative assessment and two enhancements,”
IEEE/ASME Transactions on Mechatronics, vol. 20, no. 2, pp. 944–
955, 2015.
[25] A. Domahidi, A. Zgraggen, M. Zeilinger, M. Morari, and C. Jones,
“Efficient interior point methods for multistage problems arising in
receding horizon control,” in Proc. of IEEE Conference on Decision
and Control, pp. 668–674, 2012.
[26] H. Ferreau, “qpOASES: a parametric active set algorithm for quadratic

7554
Authorized licensed use limited to: Yonsei Univ. Downloaded on July 03,2024 at 13:43:55 UTC from IEEE Xplore. Restrictions apply.