Robotic speech processing

Integrated Speech-based Perception System for User Adaptive Robot
Motion Planning in Assistive Bath Scenarios
Athanasios C. Dometios, Antigoni Tsiami, Antonis Arvanitakis, Panagiotis Giannoulis,
Xanthi S. Papageorgiou, Costas S. Tzafestas, and Petros Maragos
School of Electrical and Computer Engineering, National Technical University of Athens, Greece
{athdom,paniotis,xpapag}@mail.ntua.gr, {antsiami,ktzaf,maragos}@cs.ntua.gr
Abstract— Elderly people have augmented needs in perform-
ing bathing activities, since these tasks require body flexibility.
Our aim is to build an assistive robotic bath system, in order
to increase the independence and safety of this procedure.
Towards this end, the expertise of professional carers for
bathing sequences and appropriate motions have to be adopted,
in order to achieve natural, physical human - robot interaction.
The integration of the communication and verbal interaction
between the user and the robot during the bathing tasks is a
key issue for such a challenging assistive robotic application.
In this paper, we tackle this challenge by developing a novel
integrated real-time speech-based perception system, which will
provide the necessary assistance to the frail senior citizens. This
system can be suitable for installation and use in conventional
home or hospital bathroom space. We employ both a speech
recognition system with sub-modules to achieve a smooth and
robust human-system communication and a low cost depth
camera for end-effector motion planning. With a variety of
spoken commands, the system can be adapted to the user’s
needs and preferences. The instructed by the user washing
commands are executed by a robotic manipulator, demonstrat-
ing the progress of each task. The smooth integration of all
subsystems is accomplished by a modular and hierarchical
decision architecture organized as a Behavior Tree. The system
was experimentally tested by successful execution of scenarios
from different users with different preferences.
I. INTRODUCTION
Most advanced countries tend to be aging societies, with
the percentage of people with special needs for nursing
attention being already significant and due to grow. Health
care experts are called to support these people during the
performance of Personal Care Activities such as showering,
dressing and eating [1], [2], inducing great financial burden
both to the families and the insurance systems. During the
last years the health care robotic technology is developing
towards more flexible and adaptable robotic systems de-
signed for both in-house and clinical environments, aiming
at supporting disabled and elderly people with special needs.
There have been very interesting developments in this
field, with either static physical interaction [3]–[5], or mobile
solutions [6]–[8]. Most of these focus exclusively on a body
The first four authors contributed equally to this work.
This research work has received funding from the European Union’s
Horizon 2020 research and innovation programme under the project
“I-SUPPORT” with grant agreement No 643666 (http://www.i-support-
project.eu/).
Fig. 1: CAD design of the robotic bath system installed in a
pilot study shower room.
Fig. 2: An overview of the proposed integrated intelligent
assistive robotic system.
part e.g. the head, and support people on performing personal
care activities with rigid manipulators.
However, body care (showering or bathing) is among the
first daily life activities which incommode an elderly’s life
[1], since it is a demanding procedure in terms of effort
and body flexibility. On the other hand, soft robotic arm
technologies have already presented new ways of thinking
robot operation [9]–[11], and can be considered safe for

direct physical interaction with frail elderly people. Also,
applications with physical contact with human (such as
showering) are very demanding, since we have to deal with
curved and deformable human body parts and unexpected
body-part motion may occur during the robot’s operation.
Therefore safety and comfort issues during the washing
process should be taken into account. In addition, control
architectures of a hyper-redundant soft arm [11]–[13], in an
environment with human motion not only can give important
advantages in terms of safety but also is a challenging task,
since several control aspects should be considered such as
position, motion and path planning [14], stiffness [15], shape
[16], and force/impedance control [17].
Another important aspect of these assistive robotic systems
is the incorporation of intuitive human-robot interaction
strategies, that involve both proper and simple communic-
ation and perception of the user and modular interconnec-
tion of several subsystems constituting the robotic system.
Additionally, Human-Robot trust is a major issue for people
especially in contexts of physical interaction. As these people
experience this new environment they have difficulty to
adapt to the operational requirements. Trust is difficult to be
established if the Robot is not dependable, predictable and
its actions are perceived as transparent [18]. So the challenge
for the system to work safely and intuitively is to enhance to
basic properties of the system robustness and predictability.
In particular, we aim to establish the communication of the
elderly people with the robotic bathing system via spoken
commands. Since speech is the most natural way of human
communication, it is the most comfortable and natural way
also in human-system communication, and generally the
most intuitive way to ask for help. Also, the bathroom
scenario as well as the system used by elderly people raise
some challenges. Firstly, the necessary distance between the
speaker and the microphones imposes distant speech recog-
nition. Secondly, the nature of the environment, i.e. high re-
verberation (bathroom walls) and noise (coming mainly from
water) brings up the need for specific components like voice
activity detection, denoising, etc., that should be adapted to
the environment. Third, speech recognition for elderly people
is a challenging problem, let alone distant speech recognition,
due to their difficulty of speaking, tremor, and corruption of
speech characteristics. Last, due to the fact that the system
is destined to elderly people, it should perform robustly and
also confirm the elderly’s choices in order to avoid undesired
situations.
In this paper an integrated intelligent assistive robotic
system designed for bathing tasks is presented. The proposed
system consists of two real time perception systems, one
for audio instructions and communication and one for user
visual capturing, as depicted in Fig. 2. For this reason
we employ a speech recognition system with sub-modules
to achieve a smooth and robust human-system communic-
ation. The system can be adapted to the user’s specific
room/space, needs and preferences, since it is independent
of the microphone positions and the user can change the
set of interaction commands according to his needs. The
components of this recognition system are a Voice Activity
Detection module, a beamforming component for enhancing
and denoising, a distant speech recognition engine and a
speech synthesis module providing audio feedback to the
user. The commanded washing tasks are executed by a
robotic manipulator, demonstrating the progress of each task.
The smooth integration of all subsystems is accomplished by
a modular and hierarchical decision architecture organized as
a Behavior Tree.
II. SYSTEM DESCRIPTION
The robotic shower system, which is currently under
development Fig. 1, will support elderly and people with
mobility disabilities during showering activities, i.e. pouring
water, soaping, body part scrubbing, etc. The degree of
automation will vary according to the user’s preferences and
disability level. In Fig. 1, a CAD design of the configuration
of the system’s basic parts is presented. The robotic system
provides elderly showering abilities enhancement, whereas
the Kinect sensors are used for user perception and HRI
applications.
The robotic arms is constructed with soft materials (rubber,
silicon etc.) and is actuated with the aid of tendons and
pneumatic chambers, providing the required motion to the
three sections of the robotic arm. This configuration makes
the arms safer and more friendly for the user, since they will
generate little resistance to compressive forces, [19], [20].
Moreover, the combination of these actuation techniques
increases the dexterity of the soft arms and allows for
adjustable stiffness in each section of the robot. The end-
effector section, which will interact physically with the user,
will exhibit low stiffness achieving smoother contact, while
the base sections supporting the robotic structure will exhibit
higher stiffness values.
Visual information of the user is obtained from Kinect
depth cameras. These cameras will be mounted on the wall of
the shower room in a proper configuration in order to provide
full body view of the user, Fig. 1. It is important to mention
that information exclusively from depth measurements will
be used to protect the user’s personal information. Accurate
interpretation of the visual information is a prerequisite
for human perception algorithms (e.g. accurate body part
recognition and segmentation [21]). Furthermore, depth in-
formation will be used as a feedback for robot control
algorithms closing the loop and defining the operational
space of the robotic devices.
Finally, human-system interaction algorithms will include
distant speech recognition, [22], employing microphone ar-
rays in order to achieve robustness. This augmented cogni-
tion system will understand the user’s interactions, providing
the necessary assistance to the frail senior citizens.
III. HRI DECISION CONTROL SYSTEM
A. Behavior Tree Formulation
The Decision Control System (DCS) consists of distinct
modules organized as a Behavior Tree (BT). The basic goal
of such a system is to properly integrate and orchestrate

the sub-modules of the system, in order to achieve both
optimal performance of the system and smooth Human Robot
Interaction. The BT has two type of nodes, the Control-Flow
nodes and the Leaf nodes. The control-flow nodes are: First,
the “Selector” which executes its children until one returns
SUCCESS and can be interpreted as sequential ORs. Second
is the “Sequence” node which executes its children until
they all return SUCCESS and can be interpreted as AND. The
Leaf nodes are the ones that are exchanging information with
the external components (i.e. outside the scope of the DCS)
and are capable of changing the configuration of the system.
The formulation of a BT is described in detail in [23].
Additional nodes are the “Loop” node that repeats part
of the tree and the “Parallel” node which allows for
concurrency as in [24].
B. Robot Platform Decision Support Characteristics
Decision modules based on BT have shown to be fault
tolerant [24] because they are modular and capable of
describing complex behavior as a hierarchical collection of
elementary building blocks [25], [26]. These blocks allow
easier testing, separation of concerns per module, reusability
and the ability to compose almost arbitrary behaviors [25].
Thus a system is built that can be easily audited. Safety
auditing is required for medical devices.
The BT consists of the following submodules. First, is
the auditory module that handles the input from the speech
recognition system which enables the communication of
the user’s commands to the system. Second is the text
to speech feedback module which communicates with the
user, acknowledges the user’s commands and can request
for confirmation. Third is the robotic arm module which
according with the execution stage of the task and the
capabilities of the robot initiates, handles the changes in
the user’s requests dynamically. The robotic arm can wash
the user’s back with various patterns. For example during
scrubbing in a cyclic pattern, the user can request at any
point from the system verbally to change pattern or to stop
completely. Except for the described capabilities the user also
can request an immediate stop of the whole process for safety
reasons by issuing a “Halt” command. The user can also
increase and decrease the robot’s washing speed, and will be
able to perform additional actions as the washing system’s
capability is expanded.
C. Design of the Automation System
The system consists of modules that are easily testable by
themselves separate from the rest of the system, and func-
tionality is replicated throughout the tree due to the design’s
modularity. The pattern used is that the robot requests from
the user for a task using the speech recognition system,
gives feedback through the speaker, requests confirmation
and proceeds to execute the task based on the behavior tree.
If we need to adapt the system with new hardware then the
current functionality will be retained as long as any new
hardware conforms to the behavior tree’s protocol.
IV. DISTANT SPEECH PROCESSING SYSTEM
Our setup involves the use of multiple sensors for audio
capture. This is essential for the use cases we have employed,
because we expect far-field speech. For this reason, we em-
ploy microphone arrays aiming to also address other factors
that deteriorate performance, such as noise, reverberation and
overlap of speech with other acoustic events.
The distant speech processing system with its sub-
components is depicted in Fig. 3. After audio capture, a Voice
Activity Detection (VAD) module detects the segments that
contain human voice and it forwards this information to a
beamforming component that exploits the known speaker’s
location (on the chair) to enhance and denoise target speech.
Subsequently a distant speech recognition engine transforms
speech to text in the form of a command to the system. A
speech synthesis module is also used to feed back to the user
the recognized command in order to confirm it or not, and if
it is confirmed, the correct system prompt is spoken out to
the user, employing the same Text-To-Speech (TTS) engine.
In the next sub-sections each module is described in more
detail.
A. Voice Activity Detection
The VAD module constitutes the first step of the distant
speech processing system. Given a fast and robust VAD
module, the speech processing system can benefit in sev-
eral ways: (a) better silence modeling (less false alarms),
(b) operation of the Automatic Speech Recognition (ASR)
module in windows of variable length, (c) less computations
needed from the ASR module (operation only inside VAD
segments).
The proposed VAD system implements a multichannel
approach to determining the temporal boundaries of speech
activity in the room. The sequence of audio events (speech
or non-speech) inside a given time interval, is estimated via
the Viterbi algorithm on combined likelihood scores coming
from single-channel event models for the entire observation
sequence. These combined scores are estimated as averages
of scores based on channel-specific event models (“sum of
log-likelihoods”).
For the modeling of each event at each channel the well-
known Hidden Markov Models (HMMs) (single-state) with
Gaussian Mixture Models (GMMs) for the probability dis-
tributions are employed. The features used are baseline Mel
Frequency Cepstral Coefficients (MFCCs) with ∆’s and ∆∆’s.
The Viterbi algorithm allows the identification of the optimal
sequence of audio events for a given time interval, while the
incorporation of the multichannel score essentially leads to
a robust decision that is informed by all the microphones in
the room.
B. Beamforming
Signals recorded from a microphone array can be exploited
for speech enhancement through beamforming algorithms.
The beamformed signals can then enhance speech recogni-
tion performance due to noise and reverberation reduction.
In our specific case, noise mainly comes from water pouring,

Fig. 3: Distant speech processing system with its sub-components.
and reverberation is high, a usual case in bathroom environ-
ments. The offline system offers two beamforming choices:
The simple but fast Delay-and-Sum Beamformer (DSB)
and the more powerful Minimum Variance Distortionless
Response (MVDR) as implemented in [27]. The online
version of the system uses only DSB, because it is faster,
thus more suitable for real-time systems than MVDR.
C. Distant speech recognition
Several factors mentioned before such as noise, reverber-
ation and distance between system/robot and speaker [28]
render speech recognition a challenging task. Thus, we
employ Distant Speech Recognition (DSR) [29], [30] with
microphone arrays.
The DSR module is always-listening, namely it is able to
detect and recognize the utterances spoken by the user at
any time, among other speech and non-speech events. It is
grammar-based, enabling the speaker to communicate with
the system via a set of utterances adopted for the context
of each specific use case. For the particular use case that
will be described later, we have employed English language.
In order to improve recognition in noisy and reverberant
environments, we use DSB beamforming using the available
microphone channels.
Regarding acoustic models, initially we trained 3-state,
cross-word triphones, with 8 Gaussians per state, on stand-
ard MFCC-plus-derivatives features on Wall street Journal
(WSJ) database [31], which is a close-talk database in
English. However, mismatch between train and test condi-
tions/environment deteriorates the DSR performance. Max-
imum Likelihood Linear Regression adaptation has been
employed to transform the means of the Gaussians on the
states of the models, using data recorded in this particular
setup.
D. Speech synthesis (TTS)
A speech synthesis module is also part of the system,
and has two particular uses: First, after the DSR module
produces a result, the TTS module gives feedback to the
user, in the form of synthesized speech, about the recognition
result asking confirmation. Second, if confirmation is given
by the user (recognized by the DSR module), the recognition
result is published to the decision support system, which in
its turn selects the appropriate system prompt to return to the
speech synthesis node, in order to be played back to the user.
This way we have a two-step verification process in order to
avoid as much as possible mis-recognized commands. The
employed speech synthesis module uses the widely available
Festival Speech Synthesis System [32].
V. USER ADAPTIVE ROBOT MOTION PLANNING
The on-line end-effector path adaptation problem, of a
robotic manipulator performing washing actions over the
user’s body part (e.g. back region), in a workspace including
depth-camera, is considered. The basis of this task lies on
the calculation of the appropriate reference end-effector pose,
in order for the robotic manipulator to be able to execute
predefined washing tasks (such as pouring water on the back
region of the user) in a compliant way with the motion and
the curvature of each body part.
In addition, this approach takes into consideration the
adaptability of the system to different users. There is a great
variability to the size of each body part among different
users and each user has different needs during the bathing
sequence. Therefore, in order to compensate with different
operational conditions, the planning of the end-effector path
begins on a normalized (in terms of spatial dimensions) 2D
“Canonical” space.
Properly designed bathing paths can be followed and fitted
on the surface of the user’s body part, with use of our
motion planning algorithm, satisfying both the time and the
spacial constraints of the motion. The resulting point of the
motion path at each time segment should be transformed
from the fixed 2D “Canonical” space to the 3D actual
operational space of the robot (Task space). In particular, we
use two bijective transformations for this fitting procedure,
Fig. 4(b),(c). The first bijection is an affine transformation
(anisotropic scaling and rotation) from the “Canonical” space
to the Image space, whereas the second one is the camera
projection transformation from the Image space to the 3D
operational space, which is included in the camera’s field
of view (Task space). These bijections ensure that the path
will be followed within the body part limits and will be
adaptable to the body motions and deformations. Finally, this
motion planning algorithm takes into account the emergency
situations that might raise on an integrated system and plans
the motion of the robot from the operational position to a
safe position (away from the user) incorporating obstacle
avoidance techniques.

Fig. 4: Three different spaces described in the proposed methodology. (a) 2D space normalized in x,y dimensions notated as
“Canonical” space. The sinusoidal path from the Canonical space is transformed the rectangular area on the Image space,
using scaling and rotation. (b) 2D Image space is the actual image (of size 512 x 424 pixels) obtained from Kinect sensor.
The yellow rectangular area marks the user’s back region, as a result of a segmentation algorithm. The transformed path is
fitted on the surface of a female subject’s back region, using the camera projection transformation. (c) 3D Task Space is the
operational space of the robot. The yellow box represents a Cartesian filter, including the points on which the robot will
operate.
Fig. 5: Point Cloud representation of the user’s back region during an experimental execution. (a) Procedure Initialization
when the user ask from the system to wash its back with a cyclic motion pattern. (b) Cyclic motion execution until another
command came. (c) The robot changed its motion primitive in order to follow the user’s will to change the motion pattern
to the sinus one.
VI. EXPERIMENTS
A. Setup Description
In order to test and analyze the performance of the
integrated system, an experimental setup is used that includes
a Kinect-v2 Camera providing depth data for the back region
of a subject, as shown in Fig. 4 (c), with accuracy analyzed
in [35]. The segmentation of the subjects’ back region is
implemented, for the purposes of this experiment, by simply
applying a Cartesian filter to the Point-Cloud data. The setup
also includes a 5 DOF Katana arm by Neuronics for basic
demonstration of a bathing scenario. In this study healthy
subjects took part in the experiments for validation purposes.
The coordination and inter-connection of the subsystems
constituting the robotic bathing system is implemented in
ROS environment.
For this specific scenario, the speech processing system
setup involves a MEMS microphone array [36], which is
portable and has been found to perform equally well with
other microphone arrays, placed opposite to the user. Re-
garding distant speech recognition task, a grammar with 11
different commands has been employed. All these commands
begin with the word “Roberta”, as the keyword that activates
the system. This is justified due to the system being always-
listening: If there is no keyword, users may be confused
when addressing someone else and not the system. The set
of commands include action prompts, i.e. “Roberta wash
my back”, “Roberta scrub my back”, start/stop prompts, i.e.
“Roberta initiate”, “Roberta stop”, “Roberta halt”, type of
motion commands, i.e. “Roberta circular”, “Roberta sinus”,
velocity commands, i.e. “Roberta faster”, “Roberta slower”
and confirmation commands, i.e. “Roberta yes”, “Roberta
no”.
B. Testing Strategy
The experiments conducted include a scenario in which
the user can decide:
• the washing action from a repertoire of tasks, i.e. wash
or scrub and the region of actions, i.e. back or legs;
• the confirmation or the rejection of on action based on
dialogue in any step is necessary;

• the type of motion that the user may prefer, i.e. cyclic
or sinus motion primitive;
• to pause the action, to execute it faster or slower, to
change the executed primitive, or to emergency stop the
whole system.
Based on these options the subject is free to decide its
command in a non-linear way of execution. The robot
under the subject’s instruction is able to follow a simple
sinusoidal or a cyclic path (representing the washing and
the scrubbing action respectively), keeping simultaneously
a constant distance and perpendicular relative orientation to
the surface of the user’s respective body region. These paths
are chosen for clarity reasons for the presentation of the
experimental results.
C. Results and Discussion
Our goal is to experimentally validate that the integrated
system can handle correctly the user’s information, prefer-
ences and instructions, while at the same time it is capable
to communicate and interact successfully with the robot that
perform the bathing task.
In Fig. 5 the Point Cloud data of the user’s back region
during a successfully executed experiment is presented. After
the initialization procedure the user asks from the system to
wash its back with a cyclic motion pattern (i.e. scrubbing
or washing action). The robot took an initial position in a
constant distance from its back, as depicted in Fig. 5 (a), and
therefore the cyclic motion was executed until another com-
mand came, Fig. 5 (b). The next instruction was to change
the motion pattern to the sinus one, and the robot changed
its motion primitive in order to follow the user’s will, as it is
shown in Fig. 5 (c). The overall performance of the system
was satisfying since a full scenario was successfully executed
from different users with different preferences.
VII. CONCLUSIONS AND FUTURE WORK
This paper presents an integrated intelligent assistive ro-
botic system designed for bathing tasks. The integration
of the communication and verbal interaction between the
user and the robot during the bathing tasks is a key issue
for such a challenging assistive robotic application. We
tackle this challenge by developing the described augmented
cognition system, which will interpret the user’s preferences
and intentions, providing the necessary assistance to the
frail senior citizens. The proposed system consists of two
real time perception systems, one for audio instructions
and communication and one for user capturing. A decision
control system integrates all the subsystems in a non-linear
way with a modular and hierarchical decision architecture
organized as a Behavior Tree. All the instructed washing
tasks are executed by a robotic manipulator, demonstrating
the progress of each task. The overall performance of the
system was tested and a full scenario was successfully
executed from different users with different preferences.
For further research, we aim to ameliorate this methodo-
logy in order to meet any soft robotic manipulation special
requirements. Furthermore, the initial predefined path will be
learned by demonstration of professional carers, with the aid
of DMP approach, in order to make the bathing sequence
more human-like. The speech processing system can be en-
hanced via training with data more matched with the usecase
scenario. Additionally, more microphone arrays distributed in
space can be used in order to better enhance speech, remove
noise coming from water and mitigate reverberation which
is very high in such environments. The system is designed to
work locally but it can be adapted to work as a teleoperation
platform. Additional functionality can be incorporated and
adaptability to users with more specific impairments can
be introduced. Users that have back injuries, wounds or
other skin abnormalities can be detected by the camera, by
the carer or the user himself, towards a decision system
with the ability to produce a personalized plans. There are
planned tests with elderly patients and the feedback can be
incorporated enhancing the systems usability.
ACKNOWLEDGMENT
The authors would like to acknowledge Rafa Lopez,
Robotnik Automation, Valencia, Spain, for the CAD rep-
resentation of the robotic bath system installed in a shower
room. The authors also wish thank Petros Koutras, Isidoros
Rodomagoulakis, Nancy Zlatintsi and Georgia Chalvatzaki
members of the NTUA IRAL for helpful discussions and
collaborations.
REFERENCES
[1] D. D. Dunlop, S. L. Hughes, and L. M. Manheim, “Disability in
activities of daily living: patterns of change and a hierarchy of
disability,” American Journal of Public Health, vol. 87, pp. 378–383,
1997.
[2] S. Katz, A. Ford, R. Moskowitz, B. Jackson, and M. Jaffe, “Studies
of illness in the aged: The index of adl: a standardized measure of
biological and psychosocial function,” JAMA, vol. 185, no. 12, pp.
914–919, 1963.
[3] T. Hirose, S. Fujioka, O. Mizuno, and T. Nakamura, “Development
of hair-washing robot equipped with scrubbing fingers,” in Robotics
and Automation (ICRA), 2012 IEEE International Conference on, May
2012, pp. 1970–1975.
[4] Y. Tsumaki, T. Kon, A. Suginuma, K. Imada, A. Sekiguchi,
D. Nenchev, H. Nakano, and K. Hanada, “Development of a skin-
care robot,” in Robotics and Automation, 2008. ICRA 2008. IEEE
International Conference on, May 2008, pp. 2963–2968.
[5] M. Topping, “An overview of the development of handy 1, a rehabilit-
ation robot to assist the severely disabled,” Artificial Life and Robotics,
vol. 4, no. 4, pp. 188–192.
[6] C. Balaguer, A. Gimenez, A. Huete, A. Sabatini, M. Topping, and
G. Bolmsjo, “The mats robot: service climbing robot for personal
assistance,” Robotics Automation Magazine, IEEE, vol. 13, no. 1, pp.
51–58, March 2006.
[7] A. J. Huete, J. G. Victores, S. Martinez, A. Gimenez, and C. Balaguer,
“Personal autonomy rehabilitation in home environments by a portable
assistive robot,” IEEE Transactions on Systems, Man, and Cybernetics,
Part C (Applications and Reviews), vol. 42, no. 4, pp. 561–570, July
2012.
[8] Z. Bien, M.-J. Chung, P.-H. Chang, D.-S. Kwon, D.-J. Kim, J.-S.
Han, J.-H. Kim, D.-H. Kim, H.-S. Park, S.-H. Kang, K. Lee, and
S.-C. Lim, “Integration of a rehabilitation robotic system (kares
ii) with human-friendly man-machine interaction units,” Autonomous
Robots, vol. 16, no. 2, pp. 165–191, 2004. [Online]. Available:
http://dx.doi.org/10.1023/B:AURO.0000016864.12513.77
[9] C. Laschi, M. Cianchetti, B. Mazzolai, L. Margheri, M. Follador, and
P. Dario, “Soft robot arm inspired by the octopus,” Advanced Robotics,
vol. 26, no. 7, pp. 709–727, 2012.

[10] C. Laschi, B. Mazzolai, V. Mattoli, M. Cianchetti, and P. Dario,
“Design of a biomimetic robotic octopus arm,” Bioinspiration &
Biomimetics, vol. 4, no. 1, p. 015006, 2009.
[11] I. D. Walker, “Continuous backbone continuum robot manipulators,”
ISRN Robotics, vol. 2013, 2013.
[12] D. B. Camarillo, C. R. Carlson, and J. K. Salisbury, “Task-space
control of continuum manipulators with coupled tendon drive,” in
Experimental Robotics. Springer, 2009, pp. 271–280.
[13] F. Renda and C. Laschi, “A general mechanical model for tendon-
driven continuum manipulators,” in Robotics and Automation (ICRA),
2012 IEEE International Conference on. IEEE, 2012, pp. 3813–3818.
[14] J. Xiao and R. Vatcha, “Real-time adaptive motion planning for a
continuum manipulator,” in Intelligent Robots and Systems (IROS),
2010 IEEE/RSJ International Conference on. IEEE, 2010, pp. 5919–
5926.
[15] M. Mahvash and P. E. Dupont, “Stiffness control of a continuum
manipulator in contact with a soft environment,” in Intelligent Robots
and Systems (IROS), 2010 IEEE/RSJ International Conference on.
IEEE, 2010, pp. 863–870.
[16] B. Bardou, P. Zanne, F. Nageotte, and M. De Mathelin, “Control of a
multiple sections flexible endoscopic system,” in Intelligent Robots
and Systems (IROS), 2010 IEEE/RSJ International Conference on.
IEEE, 2010, pp. 2345–2350.
[17] A. Kapadia and I. D. Walker, “Task-space control of extensible
continuum manipulators,” in Intelligent Robots and Systems (IROS),
2011 IEEE/RSJ International Conference on. IEEE, 2011, pp. 1087–
1092.
[18] P. A. Hancock, D. R. Billings, K. E. Schaefer, J. Y. C. Chen, E. J.
de Visser, and R. Parasuraman, “A meta-analysis of factors affecting
trust in human-robot interaction,” vol. 53, no. 5, pp. 517–527.
[19] T. G. Thuruthel, E. Falotico, M. Cianchetti, F. Renda, and C. Laschi,
“Learning global inverse statics solution for a redundant soft robot,”
in Proceedings of the 13th International Conference on Informatics in
Control, Automation and Robotics, 2016, pp. 303–310.
[20] M. Manti, A. Pratesi, E. Falotico, M. Cianchetti, and C. Laschi, “Soft
assistive robot for personal care of elderly people,” in 2016 6th IEEE
International Conference on Biomedical Robotics and Biomechatron-
ics (BioRob), June 2016, pp. 833–838.
[21] A. Vedaldi, S. Mahendran, S. Tsogkas, S. Maji, R. Girshick, J. Kan-
nala, E. Rahtu, I. Kokkinos, M. Blaschko, D. Weiss, et al., “Under-
standing objects in detail with fine-grained attributes,” in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition,
2014, pp. 3622–3629.
[22] I. Rodomagoulakis, N. Kardaris, V. Pitsikalis, E. Mavroudi, A. Kat-
samanis, A. Tsiami, and P. Maragos, “Multimodal human action
recognition in assistive human-robot interaction,” in 2016 IEEE In-
ternational Conference on Acoustics, Speech and Signal Processing
(ICASSP), March 2016, pp. 2702–2706.
[23] A. Marzinotto, M. Colledanchise, C. Smith, and P. Ogren, “Towards
a unified behavior trees framework for robot control.” IEEE, pp.
5420–5427.
[24] M. Colledanchise, A. Marzinotto, D. V. Dimarogonas, and P. Oegren,
“The advantages of using behavior trees in mult-robot systems,” in
Proceedings of ISR 2016: 47st International Symposium on Robotics,
pp. 1–8.
[25] A. Klckner, “Interfacing behavior trees with the world using descrip-
tion logic.” American Institute of Aeronautics and Astronautics.
[26] M. Colledanchise and P. gren, “How behavior trees modularize hybrid
control systems and generalize sequential behavior compositions, the
subsumption architecture, and decision trees,” vol. PP, no. 99, pp. 1–
18.
[27] S. Lefkimmiatis and P. Maragos, “A generalized estimation approach
for linear and nonlinear microphone array post-filters,” Speech com-
munication, vol. 49, no. 7, pp. 657–666, 2007.
[28] C. Ishi, S. Matsuda, T. Kanda, T. Jitsuhiro, H. Ishiguro, S. Nakamura,
and N. Hagita, “A robust speech recognition system for communication
robots in noisy environments,” IEEE Trans. Robotics, vol. 24, no. 3,
pp. 759–763, 2008.
[29] M. Wölfel and J. McDonough, Distant speech recognition. John
Wiley & Sons, 2009.
[30] A. Katsamanis, I. Rodomagoulakis, G. Potamianos, P. Maragos, and
A. Tsiami, “Robust far-field spoken command recognition for home
automation combining adaptation and multichannel processing,” in
Proc. ICASSP, 2014.
[31] D. B. Paul and J. M. Baker, “The design for the wall street journal-
based csr corpus,” in Proceedings of the workshop on Speech and
Natural Language. Association for Computational Linguistics, 1992,
pp. 357–362.
[32] A. Black, P. Taylor, R. Caley, R. Clark, K. Richmond, S. King,
V. Strom, and H. Zen, “The festival speech synthesis system
version 1.4.2,” Software, Jun 2001. [Online]. Available: http:
//www.cstr.ed.ac.uk/projects/festival/
[33] Y. Zhou, M. Do, and T. Asfour, “Learning and force adaptation for
interactive actions,” in Humanoid Robots (Humanoids), 2016 IEEE-
RAS 16th International Conference on. IEEE, 2016, pp. 1129–1134.
[34] ——, “Coordinate change dynamic movement primitivesa leader-
follower approach,” in Intelligent Robots and Systems (IROS), 2016
IEEE/RSJ International Conference on. IEEE, 2016, pp. 5481–5488.
[35] L. Yang, L. Zhang, H. Dong, A. Alelaiwi, and A. El Saddik, “Eval-
uating and improving the depth accuracy of kinect for windows v2,”
IEEE Sensors Journal, vol. 15, no. 8, pp. 4275–4285, 2015.
[36] Z. Skordilis, A. Tsiami, P. Maragos, G. Potamianos, L. Spelgatti,
and R. Sannino, “Multichannel speech enhancement using mems
microphones,” in Proc. ICASSP, 2015.

Robotic speech processing

More Related Content

Similar to Robotic speech processing (20)

More from Ravindranath Tagore (11)

Recently uploaded (20)

Robotic speech processing