Shared Admittance Control for Human-Robot Co-manipulation based

on Operator Intention Estimation

Jonathan Cacace, Alberto Finzi and Vincenzo Lippiello

PRISMA Lab, Dipartimento di Ingegneria Elettrica e Tecnologie dell’Informazione,

Universit

a degli Studi di Napoli, Federico II, via Claudio 21, 80125, Naples, Italy

Keywords:

Variable Admittance Control, Human-Robot Co-manipulation, Shared Control.

Abstract:

Collaborative robots are increasingly employed in industrial workplaces, assisting human operators in decreas-

ing the weight and the repetitiveness of their activities. In this paper, we assume the presence of an operator

cooperating with a lightweight robotic arm, able to autonomously navigate its workspace, while the human

co-worker physically interacts with it leading and inﬂuencing the execution of the shared task. In this scenario,

we propose a human-robot co-manipulation method in which the autonomy of the robot is regulated according

to the operator intentions. Speciﬁcally, the operator contact forces are assessed with respect to the autonomous

motion of the robot inferring how the human motion commands diverges from the autonomous ones. This in-

formation is exploited by the system to adjust its role in the shared task, leading or following the operator

and to proactively assist him during the co-manipulation. The proposed approach has been demonstrated in

an industrial use case consisting of a human operator that interacts with a Kuka LBR iiwa arm to perform a

cooperative manipulation task. The collected results demonstrate the effectiveness of the proposed approach.

1 INTRODUCTION

The widespread of lightweight robots enables to the

development of novel service robotic applications,

where humans and robots collaborate for the execu-

tion of shared tasks (Corrales et al., 2012). Such sys-

tems, also known as cobots, can share their workspace

with human workers ensuring safe physical human-

robot cooperation and allowing humans and robots to

work side-by-side to merge their complementary abil-

ities. In this perspective, advanced human-robot col-

laborative methods can facilitate the human work in

all the operations difﬁcult to automatize both in in-

dustrial tasks, such as assembly of heavy or complex

parts (Giordano et al., 2008), and service tasks like

cyclic object manipulation in dynamic environments.

This paper presents a framework that supports

physical human-robot interaction during the execu-

tion of cooperative tasks. In the proposed approach,

the robotic system on-line adapts its behaviour ac-

cording to the operator intention, which is continu-

ously estimated from his/her contact forces. In par-

ticular, we assume a shared control system, where the

robot can autonomously execute a requested manip-

ulation action, while the human operator can physi-

cally interact with the robot end effector, adjusting or

Figure 1: The human operator physically interacts with a

lightweight manipulator during a cooperative manipulation

task.

modifying the motion of the robot or its target. In this

scenario, human interventions can be associated with

different intentions: lead the robot, slightly adjust its

motion, change the target of the action, speed up its

execution or use the manipulator as a passive tool. In

order to exploit these interaction modes in an intuitive

manner, human interventions are to be interpreted and

the robotic behaviour must be consequently adapted.

Cacace, J., Finzi, A. and Lippiello, V.

Shared Admittance Control for Human-Robot Co-manipulation based on Operator Intention Estimation.

DOI: 10.5220/0006831800610070

In Proceedings of the 15th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2018) - Volume 2, pages 61-70

ISBN: 978-989-758-321-6

For this purpose, we propose a shared control

system supporting different autonomous and semi-

autonomous operative modes, regulated according to

the operator intentions, which are interpreted consid-

ering contact forces provided by the human operator.

In this context, the contact forces are continuously as-

sessed to inform the robot about the next target point

to reach or to interpret the motion deviation intended

by the human during cooperative manipulation. The

intention recognition process relies on a 3-Layer Neu-

ral Networks that, upon receiving as input the robot

motion direction and the operator contact forces, in-

fers the intention of the operator to follow/contrast

the manipulator motion towards a target point, devi-

ate from the latter, or use the robot manipulator in

direct manual control in order to perform other ac-

tions. In this scenario, the robot can operate in ei-

ther a passive or an active mode. When the estimated

intention of the human is not aligned with the target

of the robot, the latter passively follows the human

guidance and the robot behaviour is fully compliant

to the operator contact. On the other hand, when

human interventions are coherent with respect to the

current robot target, the robotic system can keep ex-

ecuting the current task, while suitably adjusting its

motion trajectory following the corrections provided

by the operator. In summary, this work proposes a

novel mixed-initiative co-manipulation framework in

which physical interaction, interpretation of the hu-

man interventions, task/target switching, and admit-

tance adjustment are seamlessly integrated. The aim

is to conciliate, in a ﬂexible and adaptive manner, the

precision and the strength of the robotic system with

the dexterity and the decisional capability of the hu-

man operators.

In order to demonstrate the proposed framework,

we designed a human-robot interaction setup (see

Fig. 1), where the human operator cooperates with a

Kuka LBR iiwa robot during the execution of a ma-

nipulation task. In this scenario, the effectiveness of

the system is assessed by comparing the fatigue and

the effort needed by the operator to accomplish the

task with and without the assistance of the proposed

framework.

The remainder of this paper is organized as fol-

lows. In Section 2 a brief overview of related works

is presented, in Section 3 the system architecture is

described, while in Sections 4 and 6 the human in-

tention estimation process is discussed. Section 5 de-

scribes how the operator intention estimation is ex-

ploited in the shared controller. Finally, in Section 7

an experimental case study to test the effectiveness of

our approach is presented.

2 RELATED WORKS

Estimating the operator intentions in order to regulate

the robot behavior during the execution of a shared

task is crucial in any kind of Human-Robot collabo-

ration activity (Hoffman and Breazeal, 2004; Hoff-

man and Breazeal, 2007). Flexible and natural in-

teraction with humans is often needed to enable co-

operation with robots in social and industrial or ser-

vice robotic applications. This problem has been ad-

dressed by different works in the robotic literature.

For instance, a method to adapt the role of the robot

considering operator fatigue during a co-manipulation

task is presented in (Peternel et al., 2016), in which

the robot learns by imitation how to assist the oper-

ator taking contact forces into account, but without

inferring his/her intentions. More related to our work,

in (Jlassi et al., 2014) a shared trajectory generator

based on operator force contact is proposed to trans-

late human intentions into ideal trajectories the robot

should follow. In this case an on-line trajectory gen-

erator is combined with a classical impedance con-

trol system, instead we rely on an integrated inten-

tion estimation system. Other approaches exploit hu-

man intention estimation to increase the efﬁciency of

task planning algorithms ((Hoang and Low, 2013) and

(Caccavale et al., 2016)). Intention recognition meth-

ods typically consider external forces excreted by the

human operator on the robot side to regulate the low

level behaviour of the robot (Park et al., 2016)(Pe-

ternel and Babic, 2013)(Gribovskaya et al., 2011)(Li

et al., 2015). Differently, our approach is aimed to

adapt robot task execution, without inﬂuencing the

low level control of the robot.

In (Li and Ge, 2014) motion intention of the hu-

man partner is detected using the human limb model

to estimate the desired trajectory, while in (Kouris

et al., 2017) external force information is exploited to

discern between a human contact and an unexpected

collision. Human motion estimation is also deployed

in (Ge et al., 2011), where the authors exploit Neu-

ral Networks to extract human motion parameter and

predict whether the human interventions are active or

passive. In contrast, we exploit Neural Networks to

directly classify the human force contacts with respect

to the robot motion during the execution of a coopera-

tive task, as already proposed in (Cacace et al., 2018).

Several other works exploit visual sensors to pre-

dict human intentions, as in (Bascetta et al., 2011),

where human motion trajectories are monitored to

predict the human presence in robotic cells, or like

in (Cacace et al., 2016), where the authors address

the problem of implicitly selecting a robot of a team

given the sequence of commands issued by the op-

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

Figure 2: Shared control architecture.

erator. Other approaches propose probabilistic meth-

ods to infer human intentions. For instance, in (Kel-

ley et al., 2008) a Hidden Markov Model (HMM) has

been used to estimate the most likely human activ-

ity perceived by a mobile robot, or in (Awais and

Henrich, 2010) and (Best and Fitch, 2015), where

Bayesian inference and weighted probabilistic state

machines, respectively, have been used to perform in-

tention recognition.

Notice that, in contrast with other approaches to

physical human-robot interaction framework based

on human intention estimation that asses operator in-

tentions to enable a compliant physical human-robot

interaction, we use this information to switch from

an active to passive participation of the robot to the

shared task and replan when a novel intended target is

estimated from the human physical interventions. In

this perspective, the approach proposed in this work

can be related to the one proposed by (Cacace et al.,

2014) for shared teleoperation of an aerial vehicles;

however this approach is less explored in physical col-

laborative manipulation.

3 CONTROL ARCHITECTURE

In this section, we describe the proposed control ar-

chitecture, which is depicted in Figure 2. The mo-

tion of the robot is managed by the Shared Admit-

tance Controller module, whose goal is to command

the robot to reach desired targets. As for the robotic

manipulator, we assume to control the position and

orientation of its end effector, relying on the robot

inner control loop to solve inverse kinematic prob-

lem; we also assume that the (external) forces act-

ing on the gripper are directly estimated by the robot

itself. During physical human-robot interaction, the

operator contact forces exerted on the end effector

are continuously monitored by the H-L control sys-

tem that exploits the Operator Intention Estimation

sub-module to assess human intentions. This infor-

mation is exploited by the Shared Admittance Con-

troller module in order to suitably adapt the robot be-

haviour during the execution of the cooperative task.

In the following, all modules of the architecture are

further detailed.

As already stated, the Human Operator can physi-

cally interact with the manipulator moving its end ef-

fector within a deﬁned workspace. In particular, we

assume that he/she applies a force F

on the robot end

effector and perceives a force feedback F

ext

. On the

other hand, the H-L control system is responsible to

select the target point to be reached in order to ac-

complish a given task. The target point X

is sent

to the Shared Admittance Controller along with the

classiﬁed user behaviour b to start the motion of the

robot. This module is to generate the motion data X

needed to reach the target X

. In addition, exploiting

an admittance controller, this module is responsible to

combine the autonomous motion data with the con-

trol inputs F

provided by the operator. In physical

Human-Robot Interaction, the admittance control is

used to establish a dynamic relationship between the

forces applied to the robot and the displacement from

its desired position (Hogan, 1984). The admittance

controller is described by the typical second-order re-

lationship:

m ¨x + d ˙x + kx = F (1)

that can be represented as (see (Siciliano and Villani,

2000)):

(

−

) + D

(

−

) + K

− X

) = F

(2)

Where M

, D

and K

are the desired virtual iner-

tia, the virtual damping and the virtual stiffness, re-

spectively, modelling the behaviour of the system as

a mass-spring-damper system. Typically, during co-

manipulation tasks the operator aims at moving the

robotic arm in free motion. For this reason, the stiff-

ness K

is set to zero in order to nullify the elastic

Shared Admittance Control for Human-Robot Co-manipulation based on Operator Intention Estimation

behaviour of the system. The output of this module

is the compliant position X

, representing the control

command for the Position-Controlled System. Finally,

the commanded position is converted into joint values

q, which are then applied to the manipulator. When no

target is selected by the H-L control system, the posi-

tion X

is set to the current position of the end effector.

The Operator Intention Estimation sub-module is

used to asses the operator intentions. Its goal is to

estimate human intentions of following or diverging

from the autonomous motion, while classifying dif-

ferent possible behaviours of the user. The operator

intentions are inferred exploiting a Neural Network

(see Section 4) using the following information: the

contact forces F

the operator exerts on the robotic

gripper, the motion direction d

inducted by the op-

erator, and the motion direction d

generated by the

shared controller. The classiﬁed behaviour is then

sent to the Shared Controller module that uses this

information to modify its level of autonomy during

the execution of the task as described in Section 5.

In order to lead the robot towards a new position, the

H-L control system must select the target the operator

is trying to reach. For this reason, we assume that,

for each task, the system is provided with the set of

possible positions to be reached during the execution

of that task. This module exploits this knowledge to

infer the operator intention to move the manipulator

toward one of these points.

4 OPERATOR INTENTION

ESTIMATION

In this section, we detail the operator intention estima-

tion process. As already stated, this process relies on

a Neural Network classiﬁer (Bishop, 1995) properly

adapted to our domain. An Artiﬁcial Neural Network

(ANN) is composed by a list of nodes, called artiﬁcial

neurons distributed on multiple layers. Typically, data

to classify travel from the ﬁrst (input) to the last (out-

put) layer, possibly after traversing different internal

(hidden) layers. In this structure, the number of nodes

of the ﬁrst and last layer, represents the number of ex-

pected input parameters and the number of possible

output classes, respectively. In order to classify hu-

man intentions, we designed a neural network of three

layers. The input and output layers contain, respec-

tively, 3 and 4 nodes, while we considered 25 nodes

in the middle hidden layer. Our aim is to use this neu-

ral network to compare the motion direction intended

by the operator with respect to the one planned and

executed by the autonomous system in order to un-

derstand whether the human and the robot activities

Table 1: Operator behaviour interpretation.

Class ID Class Name Description

#0 Accompany The operator is

touching the end

effector without

providing any con-

tribution to the

task.

#1 Opposite The operator moves

the manipulator in

the opposite direc-

tion of the planned

path.

#2 Coinciding The operator moves

the manipulator fol-

lowing the planned

path.

#3 Coinciding deviation The operator devi-

ates from the planned

path, trying to reach a

target.

and goals are aligned or not. The input of this net-

work is calculated starting from the information about

the contact forces F

, the motion direction induced by

the operator d

, the motion generated by the controller

to follow the path toward the target point. We as-

sume that only linear segments are used to navigate

the workspace. In addition, the closest point C

the end effector with respect to the path segment un-

der execution is continuously calculated in order to

inform the system on how far is the robot with respect

to the planned trajectory. Therefore, the input of the

Neural Network is represented by:

• ||F

||: The magnitude of forces exerted by the op-

erator.

• ||X

−C

||: The distance between the current end

effector position and the closest point on the path

segment under execution.

• ∠(

−→

): The deviation between the planned and

human motions, calculated as the angle between

the two movement vectors.

Table 1 reports the classes recognized by the Neu-

ral Network along with a brief description of the as-

sociated interpretation of the operator behaviours. We

distinguish the following four behaviours (three of

them are shown in Figure 3). In the ﬁrst place, we

consider the case in which the operator is in physical

contact with the gripper, but only accompanying the

manipulator (i.e. waiting that the autonomous system

accomplishes its goal), without providing any addi-

tional contribution to the task execution (Accompany).

The second case is illustrated in Figures 3(c) and 3(d),

where the operator is driving the robot in the oppo-

site direction with respect to the trajectory and target

planned by the autonomous system (Opposite). In-

stead, in Figure 3(a), the operator is moving the robot

end effector in the same direction of the planned path

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

(a) Coinciding be-

haviour

(b) Coinciding deviation be-

haviour

Figure 3: Different operator behaviours. The red arrow represents the autonomous motion direction D

, the blue arrow is the

operator motion direction d

. The robot is moving toward the waypoint W

, while C

is the closest point between the end

effector and the planned path.

Table 2: Confusion Matrix.

0 1 2 3

0 89.2% 0.23% 10.57% 0%

1 3.2% 95.02% 0.0% 1.78%

2 6.3% 0.6% 84.1% 9.0%

3 0.0% 5.58% 5.12% 89.3%

(Coinciding). Finally, Figure 3(b) illustrates the case

in which the operator moves the end effector away

from the planned path, but he/she is still trying to

reach the planned target (Coinciding deviation).

As for the training of the Neural Network, we in-

volved a group of 10 users (students and researchers)

asking them to physically interact with the robotic

arm (10 minutes each) in different interactive super-

vised situations covering the intentions classes intro-

duced above. During the training phase, the manipu-

lator is programmed to move towards predeﬁned way-

points. This way, we collected several examples of

the operator interactive behaviour for each class. The

ﬁnal training set used for the system evaluation is

composed of about 10000 samples and the network

has been trained using the back propagation function.

Once trained, we tested the recognition system with a

different group of users similarly composed, interact-

ing with the manipulator whose task is to follow a pre-

planned squared path with its end effector. Table 2

reports the confusion matrix of the intention classiﬁer

showing that the recognition system has been able to

correctly classify 89.4% of the samples.

5 SHARED ADMITTANCE

CONTROL BASED ON

OPERATOR INTENTION

ESTIMATION

In this section, we discuss how the robot uses the esti-

mation of the operator intentions to adapt its role and

interaction mode during the execution of the shared

task. As already stated, the robot can switch from a

passive to an active operative mode. In the ﬁrst case,

the manipulator is fully compliant to operator contact

without providing any contribution to the task, while

in the second case the manipulator is actuated to bet-

ter assist the operator in reaching the target position

= (x

). Referring to schema depicted

in Figure 2, in order to reach the target state, a velocity

reference

command at a certain time i is generated

as follow:

= ω

− 2ζ

i−1

(3)

i−1

τ (4)

Where ω and ζ are gains representing frequency and

damping of the system respectively, while τ is the

sampling time of the controller. The position error

= (X

− X

) is calculated as the distance of the ma-

nipulator (X

) from the designed target position (X

The velocity obtained with the Formula 4 is succes-

sively integrated to get the desired position of the

robot:

= X

i−1

τ (5)

The compliant behaviour of the manipulator is en-

abled with the following formula, in which the com-

pliant position of the end effector is calculated con-

sidering both the operator forces and the autonomous

control data:

i+1

+ D(

−

) + F

(6)

Shared Admittance Control for Human-Robot Co-manipulation based on Operator Intention Estimation

We allow the switching from the passive to the ac-

tive mode by nullifying the autonomous contribution

of the robot from the previous equation, i.e. by set-

ting to zero the desired acceleration and velocity, as

described in Algorithm 1. In particular, we want to in-

hibit the contribution of the autonomous motion when

the user behaviour is recognized as Opposite, for this

purpose we set

= 0 in Equation 6. This way,

the robot is fully compliant to human contact forces,

without trying to came back to the initial position X

Differently, when the operator behaviour is classiﬁed

as Coinciding or Coinciding deviation we have two

different situations associated with two different tar-

get points selected by the H-L control system. In the

Coinciding case, the point to reach is represented by

the one assessed as the current operator target X

. In-

stead, in the second case, a new target point is selected

in order to help the operator in order to move back the

end effector towards to the planned path. Speciﬁcally,

in order to provide the human with a smooth guid-

ance towards the planned path, during Coinciding de-

viation, the robotic system is provided with a target

point X

, which is a midpoint between the ﬁnal tar-

get position X

and the closest point to the planned

path C

. This target point is then continuously up-

dated during the Coinciding deviation until a differ-

ent human intention is recognized. Notice that, the

described approach to Coinciding deviation not only

provides the operator with a smooth guidance toward

the planned path and the target point, but also pro-

duces a force feedback on the human side, which is

associated with a feeling of the displacement between

the current robot position and the planned path. This

haptic feedback is conceptually similar to the one al-

ready proposed by (Cacace et al., 2014) for mixed-

initiative teleoperation of aerial vehicles.

Finally, in order to avoid discontinuity of the con-

trol input that could induce instability in the system,

a smooth transition between the different levels of au-

tonomy (e.g from passive to active and viceversa or

from coinciding to deviating and viceversa) has to

be considered. For this reason we adopted a time-

vanishing smoothing term as proposed by (Lippiello

et al., 2016). Speciﬁcally, given a switch from the ac-

tive to the passive control mode which starts at time

t = 0, the velocity command introduced in Equation 7

is computed as follow:

v(t) = v

(t) + e

1/γ

(0) − v

(0)) (7)

where γ is a time constant representing the dura-

tion of the transition phase, while v

and v

are the

velocity commands related to the system acting in the

active and passive mode, respectively.

Algorithm 1: Shared admittance control based on operator

intention.

Require: Target point: X

, Behaviour: b

1: procedure X

= shared controller(X

,b)

2: while X

is not reached do

3: if b == Opposite then

= 0.0

6: else

7: e

= (X

− X

)

= ω

− 2ζ

i−1

10: if b == Coinciding deviation then

11: X

= X

12: else

13: X

= X

14: end if

15: end if

16: end while

17: end procedure

6 PREDICTION OF THE

OPERATOR TARGET

In the proposed system, the operator intention esti-

mation is also exploited to recognize his/her inten-

tion to bring the robot towards speciﬁc target posi-

tions within the workspace. In particular, we assume

that each task is associated with a set of possible states

characterized by target positions to be reached in or-

der to accomplish a given task. We also assume that

this set is available to the system and can analysed

by the H-L control system module in order to assess

the most probable target point the operator is trying to

reach. For this purpose, the system generates a set of

virtual trajectories, each connecting the current posi-

tion of the end effector with an available target point.

The operator intention is estimated with respect to all

the virtual trajectories, until the estimated intention

becomes Coinciding for only one trajectory; this tra-

jectory is then assumed as the one aimed by the oper-

ator. This decision process is exempliﬁed in Figure 4,

where the system is initialized with four states. At the

start, the operator is not providing any contribution to

the task, hence the system cannot infer the target (in

Figure 4(a) all the virtual trajectories are colored in

black). Successively, the operator starts moving the

end effector toward one of the target points, while the

system identiﬁes two of them as target candidates (in

Figure 4(b) the two red paths represent the target can-

didates). Finally, the operator intention becomes Co-

inciding for only one trajectory, hence the associated

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

(a) No target candidates

(b) Two target candidates

Figure 4: Prediction of the operator target.

state is selected as the one intended by operator and

a cartesian trajectory is generated and executed (Fig-

ure 4(c)). On the other hand, if the operator intention

is classiﬁed as Opposite during the execution of the

selected trajectory, we assume that the operator target

may be changed/reﬁned hence a novel target decision

process starts in order to generate the new target along

with the new trajectory.

Table 3: Experimental Results.

min. max. mean std.

Fatigue (S) 1.9 143.3 24.5 17.26

Fatigue (P) 1.8 202.7 45.3 30.4

Force (S) 0 32.04 9.14 8.31

Force (P) 0 57 16.59 13.9

(a) Operator fatigue in shared operative mode, compared

with the maximum and mean fatigue values in passive mode.

(b) Operator contact force in shared operative mode, com-

pared with the maximum and mean contact force values in

passive mode.

Figure 5: Operator fatigue (a) and contact force (a).

7 CASE STUDY

The effectiveness of the proposed system has been

assessed by deﬁning a collaborative manipulation

case study where a human worker cooperates with a

lightweight robotic arm in order to grasp different ob-

jects and plug them in predeﬁned locations. In order

to show the advantage of the presented approach, a

user already trained on the system performed repeti-

tive tries of the designed task. We compared the sys-

tem performance with respect to a baseline version of

the framework in which the H-L control system is dis-

abled and the manipulator operates as a completely

passive robotic assistant. We refer to this operative

mode as Passive (P). During test execution, we evalu-

ated both task performance, considering the distance

covered by the manipulator, and the human physi-

cal effort, measuring operator muscle fatigue and the

norm of the force that he/she exerts on the manipula-

tor. In this context, we measured muscle fatigue ex-

ploiting a set of electromyographic sensors, that pro-

vide us the activation level of the arm muscles. The

experimental setup is depicted in Figure 6. Tests have

been performed using the Kuka LWR iiwa robot, con-

trolled via ROS middleware running on GNU/Linux

OS, as in (Hennersperger et al., 2016). Regarding the

Shared Admittance Control for Human-Robot Co-manipulation based on Operator Intention Estimation

(a) Experimental setup (b) Grasping the ﬁrst object

Figure 6: Human operator interacting with the Kuka iiwa equipped with WEISS Wsg50 during system testing.

gripper, we used the WEISS Wsg50 controlled via a

standard joypad to close and open its ﬁngers. The

overall manipulation task consists in grasping three

objects, one at a time, bringing them to speciﬁc po-

sitions of the workspace, as shown in Figures 6(b) -

6(d). Therefore, task accomplishment requires six ac-

tions: three picks and three places. During the experi-

ment, the operator is free to decide when to interact

with the manipulator by selecting any of the avail-

able objects and positions as targets. A demonstra-

tive video is available at this link: goo.gl/7xLA4r. Ta-

ble 3 reports the mean of norm of human force and

fatigue data over all the the experiments. In this table,

the minimum, the maximum, the mean, and the stan-

dard deviation of the operator fatigue and the norm of

the force related to shared (S) and passive (P) settings

are reported. As for task performance, the mean dis-

tance covered by the manipulator end effector during

the tests is 3.6 meters in the shared setting against the

4.51 meters in the passive mode, therefore the shared

system not only enables a more comfortable interac-

tion, but also a more efﬁcient task execution.

The beneﬁts of the proposed system are also illus-

trated in Figure 5, where we compare the human force

and fatigue measured in the two settings during one of

the experiments. For this purpose, illustrate the EMG

signal (red plot in Figure 5(a)) and the force contact

one (red plot in Figure 5(b)) detected in the shared

mode (adaptive shared controller), with respect to the

maximum and mean values (upper and lower dotted

lines in Figure 5) of the same signals obtained in the

passive mode. The clear reduction of both fatigue and

force highlight the advantage of the approach.

8 CONCLUSION

We presented a framework that supports physical

human-robot interaction in collaborative manipula-

tion tasks. In the proposed approach, the human

physical interaction with the robotic system is con-

tinuously assessed in order to infer the operator in-

tentions with respect to the current robotic behavior.

The recognized intentions are then exploited by the

shared control system to on-line adjust the robotic be-

havior with respect to the human interventions. We

described the overall architecture detailing the oper-

ator intention estimation method and the associated

shared control system. The proposed system has been

demonstrated presenting the case study of a collabora-

tive manipulation tasks, where the performance of the

proposed method have been compared with respect to

a baseline system relying on an admittance controller

with disabled intention estimation and the associated

adaptive mechanisms. The collected results shows the

effectiveness of the approach in terms of both task

performance and human fatigue.

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics

ACKNOWLEDGEMENTS

The research leading to these results has been sup-

ported by the ERC AdG-320992 RoDyMan, H2020-

ICT-731590 REFILLs and MADWALK projects re-

spectively. The authors are solely responsible for its

content. It does not represent the opinion of the Euro-

pean Community and the Community is not responsi-

ble for any use that might be made of the information

contained therein.

REFERENCES

Awais, M. and Henrich, D. (2010). Human-robot collab-

oration by intention recognition using probabilistic

state machines. In 19th International Workshop on

Robotics in Alpe-Adria-Danube Region (RAAD 2010),

pages 75–80.

Bascetta, L., Ferretti, G., Rocco, P., Ard

o, H., Bruyninckx,

H., Demeester, E., and Lello, E. D. (2011). Towards

safe human-robot interaction in robotic cells: An ap-

proach based on visual tracking and intention estima-

tion. In 2011 IEEE/RSJ International Conference on

Intelligent Robots and Systems, pages 2971–2978.

Best, G. and Fitch, R. (2015). Bayesian intention infer-

ence for trajectory prediction with an unknown goal

destination. In 2015 IEEE/RSJ International Confer-

ence on Intelligent Robots and Systems (IROS), pages

5817–5823.

Bishop, C. M. (1995). Neural Networks for Pattern Recog-

nition. Oxford University Press, Inc., New York, NY,

USA.

Cacace, J., Finzi, A., and Lippiello, V. (2014). A mixed-

initiative control system for an Aerial Service Vehicle

supported by force feedback. In IEEE International

Conference on Intelligent Robots and Systems.

Cacace, J., Finzi, A., and Lippiello, V. (2016). Implicit

robot selection for human multi-robot interaction in

Search and Rescue missions. In 25th IEEE Interna-

tional Symposium on Robot and Human Interactive

Communication, RO-MAN 2016.

Cacace, J., Finzi, A., and Lippiello, V. (2018). Enhancing

human-robot collaborative task execution via contact

force classiﬁcation. Human Friendly Robotics.

Caccavale, R., Cacace, J., Fiore, M., Alami, R., and Finzi,

A. (2016). Attentional supervision of human-robot

collaborative plans. In 25th IEEE International Sym-

posium on Robot and Human Interactive Communica-

tion, RO-MAN 2016.

Corrales, J. A., Garcia Gomez, G. J., Torres, F., and

Perdereau, V. (2012). Cooperative tasks between hu-

mans and robots in industrial environments. Interna-

tional Journal of Advanced Robotic Systems, 9.

Ge, S. S., Li, Y., and He, H. (2011). Neural-network-based

human intention estimation for physical human-robot

interaction. In 2011 8th International Conference on

Ubiquitous Robots and Ambient Intelligence (URAI),

pages 390–395.

Giordano, P. R., Stemmer, A., and Arbter, K. (2008).

Robotic Assembly of Complex Planar Parts : An Ex-

perimental Evaluation. pages 22–26.

Gribovskaya, E., Kheddar, A., and Billard, A. (2011). Mo-

tion learning and adaptive impedance for robot con-

trol during physical interaction with humans. In 2011

IEEE International Conference on Robotics and Au-

tomation, pages 4326–4332.

Hennersperger, C., Fuerst, B., Virga, S., Zettinig, O., Frisch,

B., Neff, T., and Navab, N. (2016). Towards mri-based

autonomous robotic us acquisitions: A ﬁrst feasibility

study. IEEE transactions on medical imaging.

Hoang, T. N. and Low, K. H. (2013). Interactive POMDP

lite: Towards practical planning to predict and exploit

intentions for interacting with self-interested agents.

CoRR, abs/1304.5159.

Hoffman, G. and Breazeal, C. (2004). Collaboration in

human-robot teams. Proceeding of the AIAA 1st In-

telligent Systems Technical Conference, page 1.

Hoffman, G. and Breazeal, C. (2007). Effects of antic-

ipatory action on human-robot teamwork efﬁciency,

ﬂuency, and perception of team. Proceeding of the

ACM/IEEE international conference on Human-robot

interaction - HRI ’07, page 1.

Hogan, N. (1984). Impedance Control: An Approach to

Manipulation. IEEE American Control Conference,

pages 304–313.

Jlassi, S., Tliba, S., and Chitour, Y. (2014). An Online

Trajectory generator-Based Impedance control for co-

manipulation tasks. IEEE Haptics Symposium, HAP-

TICS, pages 391–396.

Kelley, R., Nicolescu, M., Tavakkoli, A., Nicolescu, M.,

King, C., and Bebis, G. (2008). Understanding human

intentions via hidden markov models in autonomous

mobile robots. In 2008 3rd ACM/IEEE International

Conference on Human-Robot Interaction (HRI), pages

367–374.

Kouris, A., Dimeas, F., and Aspragathos, N. (2017). Con-

tact distinction in human-robot cooperation with ad-

mittance control. 2016 IEEE International Confer-

ence on Systems, Man, and Cybernetics, SMC 2016 -

Conference Proceedings, pages 1951–1956.

Li, Y. and Ge, S. S. (2014). Human–Robot

Collaboration Based on Motion Intention Estima-

tion. IEEE/ASME Transactions on Mechatronics,

19(3):1007–1014.

Li, Y., Tee, K. P., Chan, W. L., Yan, R., Chua, Y., and

Limbu, D. K. (2015). Continuous Role Adaptation

for Human Robot Shared Control. IEEE Transactions

on Robotics, 31(3):672–681.

Lippiello, V., Cacace, J., Santamaria-Navarro, A., Andrade-

Cetto, J., Trujillo, M.

o., Esteves, Y. R., and Vig-

uria, A. (2016). Hybrid visual servoing with hierar-

chical task composition for aerial manipulation. IEEE

Robotics and Automation Letters, 1(1):259–266.

Park, J. S., Park, C., and Manocha, D. (2016). Intention-

aware motion planning using learning based human

motion prediction. CoRR, abs/1608.04837.

Peternel, L. and Babic, J. (2013). Learning of compliant

Shared Admittance Control for Human-Robot Co-manipulation based on Operator Intention Estimation

human-robot interaction using full-body haptic inter-

face. Advanced Robotics, 27:1003–1012.

Peternel, L., Tsagarakis, N., Caldwell, D., and Ajoudani, A.

(2016). Adaptation of robot physical behaviour to hu-

man fatigue in human-robot co-manipulation. IEEE-

RAS International Conference on Humanoid Robots,

pages 489–494.

Siciliano, B. and Villani, L. (2000). Robot Force Control.

Kluwer Academic Publishers, Norwell, MA, USA, 1st

edition.

ICINCO 2018 - 15th International Conference on Informatics in Control, Automation and Robotics