1

Motion planing and Control of Maneuverable

Human Power Augmentation Lower Limbs
Exoskeletons
Abusabah I. A. Ahmed and Samah H. H. Mohammed Khir

Abstract—The Human brain is a complex information pro- m/s, pain signals travel slower at 0.61 m/s and touch signals
cessing device, controls the human body motion and sensory travel at speeds of 76.2 m/s. If you are reading this at this
system through the nervous system and muscles. Human gait is moment and thinking at the same time, which some people
a complex Multiple Inputs Multiple Outputs (MIMO) process,
which need an accurate analysis tools to extract valuable and may have trouble with, thought signals are traveling at speeds
reliable information describing it’s dynamics. Consider mo- ranging between 20 and 30 m/s.
tion planing and control in maneuverable lower-limb power The motion control and planning of lower extremity
augmentation exoskeleton systems, the intentions identification exoskeletons have gained considerable interests in recent
through limited sensors is a great deal. Dual Reaction Force years, especially for human power augmentation applications
(DRF) sensors are developed for efficient walking speed changing
estimation. We describe the finite states for our system using [2][3][4][5][6]. The lower limb exoskeleton systems designed
Markov Decision Process (MDP). Due to different system’s state for human power augmentation, load transfer and endurance
conditions we proposed new term ”Dynamic Thresholded Reward increasing. In different applications, many control methods
(DTR)” calculated based on Recurrent Neural Network (RNN) are demonstrated for human-exoskeleton system’s navigation
to estimate appropriate system’s gait transition according to the and smooth gait transitions. The Berkeley Lower Extremity
pilot intentions. System simulations are carried out applying
fusion sensing technique. The states transition after a given action Exoskeleton (BLEEX) is the most famous exoskeleton system
in following the optimal policy thereafter to improve the system , which is actuated by hydraulic system. Sensitivity Ampli-
response in tracking. MATLAB/Simulink is used to calculate fication Control (SAC) algorithm is proposed for BLEEX
the theory of human intention and the robot motion simulation. control which aimed to reduce the sensors complexity [7].
Based on adaptive Central Pattern Generators CPGs and high SAC method efficiently controls exoskeleton robot to shad-
sensitive force sensors we develop a new control strategy for
switches between different gates (flat terrain walking, stair ascent, ow human motion but in economic point of view it’s so
stairs descent , speed up and slow down). expensive and resource consumer. In other words SAC is
expensive for both development and practical applications. Y.
Index Terms—Maneuverable Exoskeleton Systems, Admittance
Control, Dual Reaction Force (DRF) sensors, Markov Chain, Sankai et al. applied impedance control method for human
Markov Decision Process, Dynamic Thresholded Reward (DTR), enhancement version of Hybrid Assistive Limb (HAL) [8].
Re. Human Intention Estimators (HIE) are implemented for HAL
control by calculating the reference patterns of the pilot
I. I NTRODUCTION by measuring human-exoskeleton interaction directly from
Electromyography signals (EMG). H. Rui et al. proposed
We consider the natural process of walking in the human a modification for SAC, which successfully achieved and
which goes through the brain [1], nervous system, lower applied for HUman power Augmentation Lower EXoskeleton
limbs and lower limbs muscles for further development on (HUALEX) control [9]. Fuzzy-based impedance regulation for
the human exoskeleton systems control. Comparing the natural control of the coupled human-exoskeleton systems [4]. The
walking process with human made one, we try to modify the learning approach of the relationship between physical human-
minority of the current motion planing and control strategies exoskeleton interaction and dynamic factors [10]. Radial Basis
for the perfect performance. The sub-systems of the human Function Neural Network (RBFNN) designed to compensate
exoskeleton system include the main controller, signal trans- for the dynamic uncertainty error and minimize the physical
mission media, sensors and joints motors are subjected to some human-robot interaction force [11].
modifications to meet the requirements of the maneuvering. The problem of perfect tracking of a known input trajecto-
The most important specification of the natural walking system ries is a great deal in human-powered exoskeleton systems,
is the speed of a nerve impulse, which varies with the type sudden changing in direction, walking speed or gait type
of nerve impulse the nervous system is sending. Some signals during coupled human-exoskeleton system’s navigation will
such as those for muscle position, travel at speeds up to 119 lead to considerable tracking errors. Several studies were per-
Abusabah I. A. Ahmed with the Karary Center for Electronic Systems formed to investigate normal human stair ascent and descent
Technology and Consultations, Karary University, 12304 Khartoum, Sudan, [12], investigations of biomechanics and motor coordination in
abusabah22@karary.edu.sd human lower extremity during ascent and descent walking at
Samah H. H. Mohammed Khird with the College of Engineering
and Architecture, Shendi University, 00000 Shendi, Sudan, different inclinations [13][14][15]. Other investigations such as
engsamah@ush.edu.sd staircase climbing of patients with knee and hip [15][16]. The
 2

indoor applications of human exoskeleton systems required flexible connections between the exoskeleton and wearer as
special care and efficient control method, changing gait type, depicted in Fig 1. In the sagittal plane, the designed ranges
acceleration, de-acceleration, transition to stair ascent and vice of motion at the hip, knee and ankle joints are −45◦ to
versa are frequent. +45◦ , 0◦ to−135◦ and −30◦ to +30◦ , respectively. The recent
Markov Decision Processes (MDPs) are modeled as a indoors applications of human-powered exoskeleton systems
probabilistic process driven by a known Markov Chain (MC) need to pay more attention for control to obtain smooth
[17][18]. However, the performance of MDPs in the practical reference trajectories tracking. The exoskeleton links are made
is challenge due to the model parameters approximation and with ideal design (minimum weight and inertia). The Link
assumptions. As a solution of MC model parameters problem lengths are adjustable respect to various pilots. Many research-
many researches are conducted recently addressed the issue of es are conducted as a series and continuous modifications
robust performance in these decision systems [19][20]. Many for HUALEX control and performance developments. Force
researchers have studied the problem of uncertain transition sensing technology represents an important feature in human-
probabilities of MDPs , recently they have efficient perfor- exoskeleton systems for monitor the interaction between pilot
mance in uncertain state transitions applications and decision and exoskeleton, which can be used for the motion planning
making fields [21][22][23]. MDPs satisfy the Markov Chain and control of these systems. Inserting or positioning of some
property, states transitions are dependent on the actions and the sensors to analysis and estimate the interaction force between
current state. We apply finite MDPs to achieve high accuracy the pilot and the exoskeleton and between exoskeleton and
motion control of the coupled human exoskeleton system rely surrounding environment is mandatory. Because these sensors
on a policy that directly maps the intentions of the pilot to the are additional control devices for the human-robot system, it
actions to determine the next state. is important for the each sensor to be small, lightweight, and
Since Neville Hogan first introduced Impedance Controllers noninvasive. As example for daily life maneuver let’s take
(IC) [24], they have become well established specially in climbing stairs, which requires active knee extension i.e. addi-
robotics and coupled human-exoskeleton system. The main tional torque must be applied [13]. The proper interaction force
illness in IC performance is the interaction forces resulted sensors of HUALEX system make successfully investigation
during gait transitions, which lead to overshoots and under- and estimation of the pilot intentions, therefore perfect intend-
shoots in trajectory tracking. Overshoots-bounded feedback ed motions and maneuvers prediction. Beside Ground Reaction
control system performance was proved by G. Deodhare Force (GRF) sensors, Two-dimensional Interaction Force Sen-
and M. Vidyasagar, they designed a controller that achieved sors (TIFS) are developed to measure quasi-interaction force
internal stability, zero steady-state error for step input and resulting from the pilot on the exoskeleton. Since our proposed
no undershoot [32]. R. D. Hill et al. investigated a dual control strategy for gait transition is a model-based control
formulation for the problem of designing linear time-invariant strategy, the dynamic model of HUALEX project must be
controllers which minimize the convex of signal [33]. M. given.
Krstic and M. Bement demonstrated a means of obtaining M (θ)θ̈ + C(θ, θ̇)θ̇ + G(θ) = τExo + τh (1)
a non-overshooting tracking response for single input single
in which θ is the vector of each joint angle, τExo and τh
output strict-feedback nonlinear systems, they tracked arbitrary
represent the input torques from HUALEX and human wearer,
reference trajectories [34]. We applied our proposed Variable
respectively. M (θ) is the inertia matrix and a function of θ,
Admittance Controller (VAC) with bounded feedback tech-
C(θ, θ̇) is the Coriolis matrix and a function of θ and θ̇,
nique to minimize tracking error overshoots and undershoots
and G(θ) is a vector of gravitational torques. During human-
during human-powered exoskeleton system’s transitions [35].
exoskeleton system navigation τh is changing according to
The paper is organized as follows: Section II shows the
human intentions. We use a modified sensing technique to
needs for maneuverable human-exoskeleton system and the
control the motion of the human exoskeleton system according
integration of such system. We validate the performance
to the pilot intentions. In order to see the effect of sudden
illness of ordinary admittance control during gait transitions
changing in the motion trajectories, experimental trails are
in Section III. Section IV shows the local regression strategy
conducted for different obstacles avoidance considered the
application on interaction force minimization. The overshoot
relation between obstacles dimension and intention feedback
reduction technique is detailed in section V. Finally, con-
signal specifications.
clusions and some perspective on future uses and further
development of this technique drawn in section VI.
A. The effect of sensors positioning and integration
Even though you chosen the best sensor type for the speci-
II. M ANEUVERABLE H UMAN -E XOSKELETON S YSTEMS
fied application, but for the efficient performance the sensory
The sudden changing in motion trajectories for dynamic system must designed, installed, and positioned properly. The
obstacle avoidance needs modified control system’s response proper placement and functions integration of the sensory
and it’s yet new research field. The HUALEX actuated and system can cause efficient service and performance. The sensor
passively driven DoFs are designed to guarantee the shadow of system designed for optimum feedback signals extraction,
all expected pilot’s maneuvers. As a wearable exoskeleton, the every sensor serve as stand-alone and with other specified
motion range for each DOF of HUALEX is designed according sensor for the mutual feedback signal. The sensor placement
to human kinematics with some slight differences due to and cooperative performance will lead to efficient on-line
 3

The load backpack with the power

unit and main controller

Thigh and Shank semi-rigid hip and knee actuators

connections with force and encoders
sensors and inclinometers

Instrumented footboard

Fig. 1. The prototype of HUman power Augmentation Lower Exoskeleton (HUALEX).

detection of pilot intentions. The modified sensory system pathes of feedback signal and control signal are different in the
doesn’t has an extreme sensitivity, but the mutual performance real human nervous system, sensor to brain and from brain to
feature lead to the wide dynamic range of operation. The drivers. We developed the human in loop algorithm to control
prediction and estimation of the pilot intentions depends on the exoskeleton system to cover the pilot intended movements
the sensor system set of data. The theory of maneuverable efficiently. The developed algorithm will lead to meet the
coupled human exoskeleton systems has very important value requirements for accurate and efficient tracking of intended
in the emergency applications, military field and industrial pilot motions during system navigation, we try to synchronize
field. The sensed data acquisition system, and useful feedback between sensory system speed (the natural sensing of human
signal extraction method can affect the system response and pilot) and actuated joint movements.
final performance efficiency. Developing of such subsystems The intelligent sensory system developed for speed decision
can lead to good tracking for the different maneuvers in the making, the system estimates the pilot intended movements
real time. based on experimentally predefined thresholds. The developed
sensory system . The sensory system configuration is devel-
B. The effect of sensory communication speed oped considering the human to control the exoskeleton system
and cover all expected pilot intended movements efficiently.
The human nervous system uses what can be approximated
The physical path between transmitters and receivers some-
as pulse frequency modulation (PFM) to transmit information
where is guided (wired) somewhere is unguided (wireless),
through nerves. A PFM signal is a sequence of pulses of nearly
the transmission can be point-to-point, point-to-multipoint or
uniform amplitude and very short duration whose frequency
multipoint-to-point depends on the data rate and the function
carries the signals data. When pulse frequency modulating
of the receivers and transmitters. Wireless transmission system
a continuous signal, information about the original signal is
can successfully serves between sensory subsystems and main
necessarily lost due to the discretized nature of the PFM
controller hanged on backpack, and between main controller
signal; nothing is known about any changes in the continuous
and active joints drivers.
signal until the occurrence of a new pulse. PFM signals
occur in the human nervous system because of the creation
III. P ROPOSED S ENSORS S YSTEM
and propagation of action potentials [41]. However, PFM
signals are rarely used in engineering applications because A. Human-exoskeleton interaction force
they are very inefficient; much of the information contained When plan the motion of maneuverable systems for special
in a continuous signal is lost during the modulation process. missions with sudden changing in input trajectories special
Pulse frequency modulators are also highly nonlinear, and care must drawn and computer simulations must conducted
are therefore not mathematically well defined or understood. first. The intelligent sensory system, data acquisition system
PFM signals are very insensitive to noise, but this seems to and mutual feedback signal extraction system are developed
be their only positive attribute. The brain can fully process for the brain to muscles navigation in human exoskeleton
and output what our eyes see in real-time. The brain activity systems. When the interaction force changed by small amount
changes in a consistent and recognizable way when the general the input track must change by to keep minimum or with
status of the subject changes, as from relaxation to alertness in acceptable threshold. The function of our predictive al-
[42]. The brain uses massive parallel processing to perform gorithm is to determines the input needed to produce the
the equivalent of several billion operations per second. The plants desired performance (minimum interaction force). The
 4

Adaptive Central Patterns Generators (ACPGs) system starts cm tread length. The threshold for the interaction force is the
with the input signal of each controlled joint as initial state, normal range for physical human-robot interaction taken as
then a reformation made according to the pilot intended nominal value, When pilot intended to transit from flat terrain
maneuver. The conducted excremental flat walking trails, stairs walking to stair ascent (fi (t))N + 4fi (t). As the interaction
ascent, stairs descent and walking speed changing to built force increases than threshold, the designed algorithm decides
the reference values data base for intended motion decision the hight of the stairs depends on the 4fi (t). So a correction
system. The result of experimental walking trails (5 trails for to the input trajectories will adapt the exoskeleton motion to
different speed) to calculate the HCT is shown in Tab. I. The the pilot’s desired motion. The typical thresholds for the heel
resulted values indicated the proportional relation of walking contact time, interaction force during stairs ascent (170 mm
speed and HCT . According to the walking trails results shown hight). Fast sensing characteristic and high system response
achieved through proposed and developed sensing techniques.
TABLE I
T HE VALUES OF HCT FOR DIFFERENT WALKING SPEED .
B. Dual Reaction Force sensor
Number Walking Speed m/s Heel Contact Time S The new proposed Dual Reaction Force sensor (DRF) sensor
1 1 0.69 ±0.05 composed of Ground Reaction Force (GRF) sensor implanted
2 1.5 0.62 ±0.09 in the exoskeleton footboard and another one for physical
3 1.8 0.54 ±0.08 Human Robot Interaction (pHRI) implanted in pilot shoe as
4 2 0.49 ±0.04 Instrumented Shoe (IS). The walking experimental trails was
5 2.5 0.41 ± 0.07 carried out on the HUALEX to validate our control algorithm.
The wearer with a height of 181 cm and a weight of 70 kg was
asked to walk on the ground. To determine the relationship
in Tab. I, we investigate that for the HCT is vary between
between walking speed and reaction force mismatch during
0.69±0.05 and 0.41±0.07 while walking speed vary between
one gait cycle we demonstrated experimental trails, asking the
1m/s and 2.5m/s. All previous studies are considered the
pilot to speedup and slowdown within the same gait cycle. The
walking speed (m/s) as a product of stride length in meters
normal speed stance phase duration control is 0.92 ± 0.13, for
(SL)and stride frequency in seconds, and it follows that there
fast speed is 0.73 ± 0.06 as measured for normal subject as
is a wealth of information on these parameters during walking.
in [44]. We use DR sensors to identify wearer intention, this
The main walking parameters observed here are unique due to
can be extracted from the mismatch between exoskeleton’s
the novelty in sensory system. The heel contact time HCT over
heel contact time ExoCT and human’s heel contact time hCT
different walking speed, provided in Tab. I showed temporal
within one gait cycle.
gait data for specific pilot. For the future application of such
We consider walking case, the range of considered speeds
sensory system some measures will be considered in our
is from 1m/s to 3m/s. The expected DR sensors signals
future work for adjustable, the extracted feedback features
deviation or mismatch according to pilot intention can be
must be able to correspondingly adjust with respect to various
lead or lag, taking exoskeleton footboard vGRF signal as a
pilot length. The HCT decreases approximately linearly with
reference in case of acceleration intention the peak of shoe
walking speed. The walking trails were carried out at Center
vGRF sensor will lead the peak of footboard FRF sensor
for Robotics, School of Automation, University of Electronic
and vice versa in slow down case as depicted in Fig. 2. The
Science and Technology of China (Chengdu, China). In these
shape of the reaction forces is typical of what is reported
trails, the selected pilot wears HUALEX and walks naturally
in the literature for full walking step, the shape shown in
according to the given speeds as depicted in Tab. I the
Fig. 2 is just for heel strike and it’s considered for mismatch
resulted HCT are registered. The walking speeds range of
detection subsequences. This force has the characteristic of
[1m/s 2.5m/s] is wide enough to evaluate the effectiveness
double hump. The first is related to weight acceptance when
of proposed methodology.
the body’s downward velocity is being arrested as shown in
For the stair ascent or stairs descent cases, the interaction
Fig. 2, the second hump is due to push-off and demonstrates
forces resulted from pilot intentions are collected from the
that the body’s centre of mass is being accelerated upwards to
effective wearable sensors in thigh and shank straps (instru-
increase its upward velocity.
mented straps). The designed force sensor is omni-directional,
to monitor and measure the rear and front interaction forces.
IV. M ARKOV D ECISION P ROCESSES
When the pilot intens to switch from flat terrain walking to
stairs ascent the feedback collected from the front force sensor, Markov Chain (MC) models [46] are used in situations in
for the vice versa action collected from the rear force sensor. which there are a large number of objects which can be in
The thigh and shank interaction forces thresholds for the gait any of several states Si , i = 1, 2, ..., n (or conditions) and
transitions are empirically designated. which move between these. The powerful tool for planning
A higher stair will require you to bend your knees more and classification in the systems transcended by uncertainties
deeply and the greater the amount of knee flexion. The joints presence is Markov Decision Process (MDPs). MDPs defined
flexion for stair ascent are adapted from the work in [43]. through the following objects:
The interaction forces on thigh and shank investigated from • Finite state space Si , i = 1, 2, ..., n
n×n
excremental trails for the stairs with 17cm step height and 28.5 • The Dynamic Uncertainties Matrix λi,j , Λ
 5

80
Exo Footboard
Instrumented Shoe
Vertical Ground Reaction Force (%BW) 70


60 Stairs Ascent

50  S1
1,0
40  0 ,1
30
 0 ,4  0 ,3
20 Slow Down Flat Walking Speed up

10
S4 S0 S3
0
0 10 20 30 40 50 60 70 80 90 100 2,0 0,2
Gait Cycle %

Fig. 2. The walking speed control sensors system feedback signals expressed
pilot intention for speedup.
 Stairs Descent 
S2
• Transition probabilities P (Si /Sj ; λi,j )
The key feature is that there is fixed probability Pi,j (tran- Fig. 3. The Markov Chain representation.
sition probability) of moving from any state to any other
state as shown in Fig. 3. The number of states is finite
by the human exoskeleton system operation conditions and low cost monitoring and controlling. SNs require adaptive and
the surface terrain of the operational area and depends on robust methods to address feedback signals overlap, mutual
system applications. The transition probability does not change performance and resource optimization. For the simulation
over time and does not depend on any previous states. The purpose the sensors number is finite and limited (si , i =
probability of different switches or transition depends on 1, 2, 3, 4), the observed quantities are also limited and inde-
the Dynamic Uncertainties Matrix (DUM) let’s define it as pendent of sensors number. The sensors positions according to
λi,j ∈ Λn×n . By default, the set of emissions (system states) the exoskeleton joints have a great sense in the pilot intentions
is n, where n is the number of possible emissions, but you estimation, this mean the designed algorithm is valid for
can choose a different set of numbers or symbols. The system the predefined pilot body dimensions. s1 and s2 are omni-
differential equation is shown in Eqn. 2. directional, so they can give a feedback for turnings left and
right, stairs ascent and stairs descent of pilot intentions. Right
n    k 
 dPS (t)  
X
λk,1 PSn (t)  
X
λ1,n  
now the HUALEX has two DoFs in the hip joint, so we
1 PS1 (t)
 dt 
 i=1

 
  i=1

 consider stairs ascent and descent cases. Excremental walking
.  .  
− .  . 
= . . (2)
trails are conducted on stairs ascent and descent to determine
 
. . 
.  n . .

    
  . 


dPS (t) n
n
 X
λk,n PSn (t)
  X  PSn (t) the relationship between joint angles and human-exoskeleton
dt λn,n
   
i=1 i=1 interaction force during stairs ascent and descent. As reported
Let P be a nxn matrix with coordinates {Pi,j : i, j = in the literature, the connections between exoskeleton and the
1, 2, ..., n}. A random process (X0 , X1 , ...) with finite state pilot near to the active joints through thigh and shank cuffs can
space S = {S1 , S2 , ..., Sn } is said to be a homogeneous measure and monitor the interaction force resulted from the
Markov chain with transition matrix P, if for k all n, all pilot intentions. According to mainly flexion, extension, turn-
i, j ∈ {1, 2, ..., n} and all i0 , ..., in−1 ∈ {1, 2, ..., n} we have: ing right or left Omni-dimensional interaction force sensors
(OIFS) are developed to measure the interaction force resulting
P (Xn+1 = Sj |X0 = Si0 , X1 = Si1 , ..., Xn = Si ) = from the pilot on the exoskeleton. The interaction forces acting
(3)
P (Xn+1 = Sj |Xn = Si ) = Pi,j between exoskeleton and pilot’s body are used as an input of
The future is independent of the past given the present and the data classifier system to predict and decide the next state
the conditions in the Eqn. 4 must satisfied. transition. The timely intention signal classification depends on
n
the current system state and the Dynamic Uncertainties Matrix
Pi,j ≥ 0,
X
Pi,j = 1 f orall i (DUM) value. The transition process sometimes subjected to
(4)
j=1
two phases, as example for transition from flat terrain walking
to stair ascent depending on the 4fi (t) stairs height will be
V. S EQUENTIAL D ECISION M AKING P ROCESS estimated.
The performance of DRF sensors for continuous plantar-
A. Sensors Regrouping force measurement and gait-phase detection during walking
Sensor Networks (SNs) consist of limited number os sensors can produce useful feedback signals for efficient and high re-
mutually worked on-line to monitor and control specified sponse walking speed control. The monitoring of such signals
dynamic process. These sensors cooperate for effective and during experimental walking trails gave us an inspiration of
 6

this control methodology. Results show that the system owns

high accuracy and response speed. The mutual performance of s1 1,2
identified sensors appear on acceleration deciding procedure s2
Neural Network
Predictors  2 ,1
depends on the feedback signals difference, from s3 and s4 Sensing
Terminal
we captured heel strike and heel off to count the difference Dynamic m,m
sn Thresholds
in stance time 4HCT . Samples of accepted thresholds during
flat terrain walking are shown in Fig. 4, from excremental
walking trails ±0.13S is investigated to be normal 4HCT .
Also as mentioned above the walking speed acceleration or Fig. 5. The block diagram of pilot intention classifier, n represents the sensors
number, m is the states number.
slow down subjected the the initial walking speed changing
and the changing value depending on the value and the sign
of 4HCT . The 4HCT is the defined as the difference in the The expected gait transitions determine the DUM elements
system heel rise phase, when pilot intends to rise his heel or which must be calculated every time step, so we can eliminate
delayed this process During walking. the non-expected λ0 s and put them to zero always.

B. Intention Classification Procedure VI. S TATES TRANSITION MANAGEMENT

The real time pilot intentions detection and classification is The transitions between system’s states is stochastic process
important task in the motion control of maneuverable coupled controlled by DUM elements which depend on the pilot inten-
human exoskeleton systems. The accurate classification is tions assume error free sensors’ performance. The thresholds
determination process represents an important issue for the
perfect pilot’s intentions estimation. The only drawback of our
Human Shoe
Exo Footboard control methodology is dependance on the pilot body mass and
dimensions. The mentioned above thresholds are determined
by walking trails are conducted by the pilot with a height of
181 cm, weight of 70 kg and 42 British shoe size.

A. Dynamic Thresholded Reward (DTR)

The DUM changes at any arbitrary time according to pilot
intentions to contain nonzero elements depending on pilot in-
tentions which indicate the system going to execute acceptable
transition with a maneuver. Some maneuvers or motions are
mutual, so the DUM elements may contain more than one
Heel Contact Time
non-zero elements. In this case the execution order decided
Fig. 4. The filtered feedback signals of heel contact time when the pilot asked by the transition priority algorithm according to the current
to make no-intention walking trails. state next transition states. The general framework of proposed
control system contains a few assumptions such as the sensors
still one of the most important issue in the data mining system is accurate and delay time free, DUM elements values
field. Many researches are conducted to design classification represents the Dynamic Thresholded Reward (DTR) and they
algorithms and efficient classifiers systems for large data sets. are a functions of t and the intended transition achieved
Multi sensors network technology is the beneficial solution for immediately to allow the system consider the future state as
information sensing and useful data collection in complicated a current state with in the given time slot 4t. Prediction of
systems [53]. The measured parameters include interaction the next move based on historical data or the finding of the
force value fi (t) at the thigh and shank points, the heel contact optimal transition in each state is a random variable depends
time HCT of pilot shoe and exoskeleton footboard. In this on the next assumptions:
work for simplification we considered 4 sensors to measure 4
• Sensors efficiency is pretty good to handle the pilot’s
variable parameters, the dynamic thresholds are updated online
intensions.
according to the current state. The identified thresholds are
• Free delay time between the decision maker nodes and
just valid for the current pilot and they need to be calibrated
joint controllers.
for different exoskeleton user. The block diagram of dynamic
• The non-acceptable transitions corresponding DUM’ el-
thresholds based intention classification system is shown in
ements always put to zero.
Fig. 5. The DUM is simplified according to the considered
• The time interval 4t is sufficient for observations pro-
gait states, the initial DUM is depicted as in 5.
cessing and transitions when needed.
 
λ0,0 λ0,1 · · · λ0,4
λ1,0 λ1,1 · · · λ1,4  B. Accessibility and Communicating States
λi,j (t) =  . (5)
 
.. .. .. 
 .. . . .  A state j is said to be accessible from state i if there is a
λ4,0 λ4,1 ··· λ4,4 positive-probability path from i to j , that is, if there is a finite
 7

sequence of states k0 , k1 , ..., km such that k0 = i , km = j need to manage the execution of tasks (transitions) to satisfy
, and P (kt , kt+1 ) > 0 for each t = 0, 1, ..., m − 1. States the efficiency, safety and robustness of navigation. Priority
i and j are said to communicate if each is accessible from rule-based method (PRBM) rely on the reception of more
the other. This relation is denoted by i ↔ j. Communication than one feedback signal on the main controller to achieve
is an equivalence relation. In particular, it is transitive if i intended transitions can lead to efficient navigation [52]. This
communicates with j and j communicates with k then i work shows how system performance decreases as number of
communicates with k. In other words, for our system shown states is incremented and transitions are managed well. The
in Fig. 3 from flat terrain walking can transit to acceleration execution of the high priority transition to be achieved first,
then to stair ascent, all these states are accessible from each then the low priority task will take the turn of execution. The
other and any two of them are communicated. The task or states priority decided to be dependent on current state and
movement must be done as a response of the system to the intended transitions. As example from speedy flat walking one
sensors output according to the pilot intention, the response must slowdown and go upstairs, the feedback signals for those
must be bounded to some thresholds to satisfy the smooth and maneuvers are from separate sensors so the DUM will has two
harmony in the system’s navigation. From the simple structure non-zero elements.
of a control system shown in Fig. 6, we’re going to arrange
or schedule the tasks or state transitions to get an optimum VII. M ODEL F ORMULATION
performance. In this section, we are going to describe the whole system
components for efficient transition mechanisms consider the
model in which the system states are finite for simulation
simplification. Furthermore, the actions available at any state
are organized and controlled according to the DUM at any
Output Actuators time during system’s navigation. The whole proposed system
configuration is designed considering the following protocol
for controlling the gait transitions of exoskeleton system:
1. The system states are finite and the transitions are limited.
CPU Input Sensors 2. The expected DUM elements are calculated every time slot
according to the pilot intentions.
3. The thresholds of the interaction forces are dynamic de-
pendant on the current state.
Controller Memory
4. The time slot is chosen considering the maneuverability of
the system and calculation complexity.
5. The stair ascent case is without height specification (in the
future we will estimate stair’s height)
6. The acceleration and slowdown are exactly estimated, so
Fig. 6. The Simple structure of an arbitrary control system. the move and the changing value.

A task is an executable software entity specialized to perfor-

A. Single hidden layer Neural Networks (SNNs)
m a special transition (walking speed acceleration, Stair ascent,
Turning right ,...), which includes feedback signal and relative In this project for the DUM elements identification we
data and computer resource. The sensory system decided by use Single hidden layer Neural Networks (SNNs), to fit the
the system application and navigation environment, in this dynamical input-output relationship we set neurons in the
paper we take the feed back signals as Ψi (t)i = 1, 2, ..., N . hidden layer to 10. The variety of sensors in our sensing
The sensed signals, useful feedback signals and sensors are system amended us to divide decision making stage in two
not dependent, this mean we applied intelligent extraction sections. So we design Single hidden layer Recurrent Neural
technique to get more signal from the limited number os Network (SRNN) which feedback connections, these kind of
sensors and we extract more feedback signals from given networks are considered more suitable for dynamical sys-
sensed signals. At any time t with the current state S0 (t) the tems [54]. The second section contains Single hidden layer
−−−→ Feedforward Neural Network (SFNN). The designed SNNs
Ψi (t) decides the next state of the human exoskeleton system
model can make predictions on the DUM elements every time
Si (t + 4t).
slot 4t. The hidden layer of the designed HNNs is the key
component because of the neurons special output function
C. Transition Priority which can do major calculations and produce the output. Each
During the system navigation, the feedback signals from neuron takes a set of input data (specified sensors); each is
the sensory system will define the next transition according to associated with a weight and a numerical value known as
pilot intentions. Experimentally, we validate that sometimes bias. The output of each neuron is a function of the output
the controller can receives two transitions in same time as of the weighted sum of each input plus the bias. Let’s define
`
example the pilot intends to turn right and accelerate the the used notations, weights are ωj,k , biases b`j , and activations
`
current walking speed or stop and sit down. These cases aj where ` is for the layer and j is for the neuron. For the
 8

transition control of exoskeleton the VAC algorithm [35] is not 4t, for the vector of inputs si (t), i = 1, 2 the SRNN can be
enough to ensure tracking error. The mal-performance is due described as:
to the considered interaction forces convex during transitions.
For this reason, we design the HNN controller such that it 1
h(t) = σ((si (t) − si (t − ∆t)) ∗ wi,j (t)) + b1i (6)
takes the decision making part on which the VAC algorithm
has shown its limitation. The recurrent hidden layer contains
2
one stage which has zero-values weights, the input of non- λi,j (t) = f (σ, wi,j ) + b2i (7)
effective neuron to the next layer. Here we mean the sensor
si which has no affection on the specified uncertainty λi,j . At each time step 4t, the input vector si (t), i = 1, 2 along
1
with the connecting weight vector wi,j (t), j = 1, 2, are inputs
to the hidden layer to produce a sequence of the hidden state
B. Pilot Intensions Identification h(t), b1i and b2i are the bias of hidden layer and output layer.
1) Single hidden layer Recurrent Neural Network (SRN- Then a prediction output λi,j (t) is learned through the hidden
N): The Recurrent Neural Networks (RNNs) used for data layer sequence and weight vector connecting the hidden layer
classification, considered the past and current data to estimate to the output layer. The iteration equations parameters are
the future values analogous to the human brain. By adjusting estimated via the on line estimation by providing a set of
the weights of an artificial neuron we can obtain the desired
output for specific input values. The feedback loop is a
delay unit with a unity weight. To validate the efficiency 11,1  21,1
b11
of the proposed transitions control methodology we limited
s1  2 2,1 b 21 i , j
the number of states for sake of the simplicity, but in the 110,1

real life the states number is much bigger depends on the

maneuverable exoskeleton application field. The simplicity of
11,2
  21,10
s2 110,2  2 2,10 b22 i , j
sample states allow the hand finding of the necessary weights,
1
while it would be quite complicated for the real life system. b 10
Inputs Output Layer
Adjusting algorithms is needed for the big number of states
Hidden Recurrent
to find the necessary weights of the RNN in order to obtain
Layer
the desired outputs of specified inputs. The weights finding
process is called learning or training of the RNN. The most
known Recurrent Neural Network (RNN) model, Elman neural Fig. 7. The proposed structure of predictive Single hidden layer Recurrent
network [54] is used for gait changing identification. The Neural Networks.
typical Elman RNN have one hidden layer with time delayed
inputs and desired outputs. In this work the SRNN acts as an
feedback. The simple structure of the NN is proportional to
observer which allows the estimation of the DUM elements to
system complexity, the limited number of states taken for this
control the state transition in MDP. The dynamic thresholds for
work is simple and can be presented with one hidden layer
the interaction force sensors (s1 and s2 ) for pilot’s intention
see Fig. 7.
classification are depend on the current state of the system.
In this part of the work we are going to implement simple
Hence, the weights are dynamic according to the input data
Single hidden layer Recurrent Neural Network (SRNN) for
set the learning rule in Eqn. 8 means that the weight vector
the learning of the pilot intensions for gait type changing. The
changing is automatically invistigated. Depends on the current
stairs case estimation depends on the past value of the inputs,
state the network weights are updated iteratively as follows:
every time step the output λi,j (t) is a function of si (t) and
si (t − 4t). The desired output λi,j (t) for transitions from wi,j (t + 4t) = wi,j (t)) + αφ[wi,j (t), si (t), di (t)]si (t)
flat terrain walking to stairs ascent is +1, and for the vise (8)
versa is -1. All non-zero values mean the pilot intends to where wi,j are the network’s weights, α is the learning rate
make some maneuver, the value identify the maneuver type. parameter and di is an error.
The zero weights between the inputs and the hidden layer 2) Single hidden layer Feedforward Neural Networks
indicate that the future output of this route is independent (SFNNs): Feedforward Neural Networks (FNNs) are artificial
of these inputs. Every 4t according to the current state and neural networks consisting of multiple layers of neurons [55],
the pilot intentions the SRNN model to learn the estimation [56]. Activation functions are typically not calculated in the
process of the DUM elements. Consider the current state the input layer, which simply passes the the inputs through to the
network calculates all the output nodes every time slot 4t, hidden layer. FNNs have been effectively applied in learning
the current state is stochastic during system navigation. As and feature selection for probabilistic output as in [57]. Single
example when the system in in S0 flat terrain walking the hidden layer Feedforward Neural Networks (SHNNs) consid-
non-zero weights for the hidden layer to the output layer. The ered the pilot intentions can estimate the the speed changing
inputs to the hidden layer from s1 and s2 every time slot 4t output depends on the current inputs of two sensors s3 and
are s1 (t), s1 (t − 4t), s2 (t) and s2 (t − 4t). Given a sequence s4 . In the construction of SFNNs shown in Fig. 8. For the
of inputs si (t), i = 1, 2 if the current state is S0 the network SFNNs the learning process through supervised training can
computes the predictions λ0,j , j = 1, 2, 3. At every time slot be described as in Eqn. 9 and 10.
 9

connections to zero after every training epoch, so it is initial

1
h(t) = σ((s1 (t) − s1 (t)) ∗ wi,j (t)) + b1i (9) value is always 0. The training set is taken from excremental
walking trails, the interaction forces (s1 and s2 ) and the
2 desired output corresponded to the intended gait λi,j .
λi,j (t) = η(f (σ, wi,j ) + b2i (10)
Simulation experiments were carried out in order to evaluate
The special sigmoidal function is designed to allow larger the performance of both NN type with respect to the given
numbers later, so it is not clipped at the top [0, 1  so the sensors input si (t), i = 1, 2, 3, 4, and the desired values of
upper value can go until (+1.6) when needed. The changing λi,j (t). The performance of SRNNs in the pilot intentions
of the walking speed during system navigation depends on estimation is shown in the Fig 9. The following parameters
the current walking speed for the pilot’s safety, the minimum were set for the SRNN training: sampling time is 0.02s,
walking speed factor is 0.4 and the maximum is 1.6. According learning rate =0.3, momentum =0.6, error tolerance is 0.001.
the mentioned above requirements, the accepted output of the
SFNNs output due to the scaled activation function is:
Training data
1. 1.6 ≥ λi,j (t) > 1 for walking speed acceleration 2.5 SRNN output
2. λi,j (t)=1 for no transition
3. 1 > λi,j (t) ≥ 0.4 for slowdown 2

1.5

Lambda(i,j) value
11,1 b11  21,1 1

s1  2 2,1 b 21 i , j
110,1 0.5

 1
1,2
  21,10 0
s2  1  2 2,10 b22 i , j
10,2

b110 −0.5
Inputs
Output Layer
Hidden Layer −1

0 0.5 1 1.5 2 2.5 3 3.5 4

Time

Fig. 8. The proposed structure of predictive Single hidden layer Feedforward

Neural Networks. Fig. 9. The performance of predictive SRNN, the case of mutual intention is
estimated perfectly (pilot intends to go upstairs and downstairs).

The performance of SFNNs in the pilot intentions estimation

VIII. E XPERIMENTAL R ESULTS
is shown in the Fig 10. The following parameters were set for
The simplicity of the learning function allows the con- the SFNN training: sampling time is 0.02s, learning rate =0.2,
struction of simple SRNN with one hidden layer contains 10 momentum =0.6, error tolerance is 0.001. The performance
neurons. We suppose that the initial state of the system is of the proposed maneuverable exoskeleton system navigation
flat walking speed S0 see Fig 3, the initial weight values control compared to our previous work the Variable Admit-
and biases are need to be identified, which is called the tance Control [35] to validate the transition efficiency on
learning process. In other words, learning is the process to interaction force minimization. For the gait type change the
determine the weights and biases of the SRNN model. The case of changing gait from flat terrain walking to 200mm
system parameters are estimated and corrected online through height stairs is simulated, result is shown in Fig 11. The second
supervised learning algorithm by provide a set of inputs case is walking speed acceleration, the speed is changed from
data for different interaction situation and its corresponding 1.5m/s to 1.8m/s result is shown in Fig. 12.
target outputs. The designed activation functions are different The classification results of the given number of sensors
according to the targeted output, sigmoid activation function were obtained using 2.3 GHz, Core I5 CPU with 4 GB of
for walking speed decisions and step function for gait type RAM, Windows 7 and Matlab 7.0 software tools.
transitions (+1 for stair ascent or flat terrain depend on the
current state, 0 for no transition and −1 for stairs descent or
flat terrain walking depend on the current state). The desired IX. C ONCLUSIONS AND FUTURE WORKS
output was normalized for the SFNN according to the intended In this paper, we proposed a technique for on-line iden-
transition in the case of walking speed changing λi,j (t) with tification of parameter subsets based on sensory data and
in the interval [0.4−1.6]. The training procedure of the SRNN apply them to a MC model predicting the next step of the
setting the recurrent connections as first-order tap delay lines, system. Applying RNN to minimize the time delay in the
which store the current output action hi (t) of the hidden layer sensory signals transmission for optimum motion recognition
and feed them back to be part of the input for calculating the and control during human exoskeleton system navigation.
layer output for the next time step hi (t + 4t). The adaptive Our control method minimized the tracking error during
algorithm of SRNN sets the values stored in the recurrent gait transition for human-exoskeleton system by reducing the
 10

30
Training data VAC
2.5 SFNN output Proposed Strategy
20
2

10
1.5

Interaction Force (N)

Lambda(i,j) value

0
1

0.5 −10

0 −20

−0.5
−30

−1
−40
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time Time (S)

Fig. 10. The performance of SFNN, the case of mutual intention is estimated Fig. 12. The performance of SFNN, the case of mutual intention is estimated
perfectly (pilot intends to change walking speed). perfectly (pilot intends to change walking speed).

15
VAC state network.
Proposed Strategy
In the future work will focus on the system response
10 optimization to minimize the transition error, and transition
time towards form brain to muscles, then HUALEX can moves
5 freely for more complicated missions. Also we’re planning
Interaction Force (N)

to generalize the thresholds identification according to the

pilot’s length and foot size and adapt to different stairs height
0
and tread length this will overcome the proposed technique
limitations.
−5

R EFERENCES
−10

[1] S. A. Zeki, “ A Vision of the Brain”, Oxford Blackwell, 1993.

[2] H. Kazerooni, A. Chu, R. Steger, That Which Does Not Stabilize, Will
−15 Only Make Us Stronger, International Journal of Robotics Research, vol.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (S) 26, 2007, pp 75-89.
[3] A. Zoss, H. Kazerooni, A. Chu, “On the Mechanical Design of the Berke-
Fig. 11. The performance of SFNN, the case of mutual intention is estimated ley Lower Extremity Exoskeleton (BLEEX)”, International Conference
perfectly (pilot intends to change walking speed). on Intelligent Robots and Systems, 2005, pp 3132-3139.
[4] H.T. Tran, H. Cheng, M. K. Duong, H. Zheng, “Fuuzy-based Impedance
Regulation for Control of the Coupled Human-Exoskeleton System”,
IEEE International Conference on Robotics and Biomimetics, 2014, pp
interaction force overshoot. The human intentions on Human- 986 - 992.
[5] Y. Sankai, “HAL: Hybrid Assistive Limb based on Cybernics”, Robotics
powered Exoskeleton can represents the Exoskeleton brain for Research, 2011, pp 25-34.
autonomous navigation and high level control performance. [6] S. Lee, Y. Sankai, “Power Assist Control for Walking Aid with HAL-
The proposed feedback signal smoothing scheme for online 3 Based on EMG and Impedance Adjustment around Knee Joint”,
gait transition control system with limited gait modes (flat International Conference on Intelligent Robots and Systems, 2002, pp
1499-1504.
terrain walking to stairs ascent) will find a more attentions [7] H. Kazerooni, J.L Racine, L. Huang, R. Steger, “On the Control of
in future developments forwards fully autonomous HUALEX the Berkeley Lower Extremity Exoskeleton (BLEEX)”, International
robotic system. As shown in the results the limitation of the Conference of Robotics and Automation, 2005, pp 4353-4360.
[8] S. Lee, Y. Sankai, “Virtual Impedance Adjustment in Unconstrained
proposed transitions control are related to the thresholds which Motion for an Exoskeletal Robot Assisting the Lower Limb”, Advanced
are dependant on the pilot’s height and foot size. Robotics, vol. 19, no. 7, 2005, pp 773-795.
Feedforward neural networks have been widely studied and [9] R. Huang, H. Cheng, Q. Chen, H. T. Tran, X. Lin, “Interactive Learn-
ing for Sensitivity Factors of a Human-powered Augmentation Lower
used in many applications, while the recurrent neural networks Exoskeleton”, IEEE/RSJ International Conference on Intelligent Robots
are still not fully exploited because of its tedious training and and Systems, 2015, pp 6409-6415.
complex system structure. We have developed a new recurrent [10] H.T. Tran, H. Cheng, X. Lin, R. Huang, “The Relationship between
Physical Human-Exoskeleton Interaction and Dynamic Factors: Using a
neural network toolbox, namely the RNN-Tool, based on two Learning Approach for Control Applications”, Science China Information
RNN structures, the recurrent multilayer perceptron and echo Science, vol. 57, no. 12, 2014.
 11

[11] M. Ka, H. Cheng, T.H. Toan, Q. Jing, “Minimizing Human-Exoskeleton [38] J. Mattingley, S. Boyd, “Real Time Convex Optimization in Signal
Interaction Force Using Compensation for Dynamic Uncertainty Error Processing”, IEEE Signal Processing Magazine, 2010, Vol. 27 , no. 3,
with Adaptive RBF Network”, Journal of Intelligent and Robotic Systems, pp. 50-61.
2015, pp 1-21. [39] C. D. Hoover, G. D. Fulk, K. B. Fite,“Stair Ascent With a Powered
[12] R. Riener, M. Rabuffetti, C. Frigo, “ Stair Ascent and Descent at Transfemoral Prosthesis Under Direct Myoelectric Control”, IEEE/ASME
Different Inclinations”, Gait and Posture, vol. 15, 2002, pp. 32-44. TRANSACTIONS ON MECHATRONICS, vol. 18, no. 3, 2013, pp 1191-
[13] C. D. Hoover, G. D. Fulk, K. B. Fite, “Stair Ascent With a Powered 1200.
Transfemoral Prosthesis Under Direct Myoelectric Control”, IEEE/ASME [40] J. E. Colgate, G. A. Ollinger, M. A. Peshkin, A. Goswami, “A 1-
Transaction on Mechatronics, vol. 18, no. 3, 2013, pp. 1191-1200. DOF Assistive Exoskeleton with Virtual Negative Damping: Effects on
[14] A. N. Amirudina, S. Parasuramanb, A. K. AhmedKhanc, I. Elam- the Kinematic Response of the Lower Limbs”, IEEE/RSJ International
vazuthid, “ Biomechanics of Hip, Knee and Ankle joint loading during Conference on Intelligent Robots and Systems, 2007, pp 1938-1944.
ascent and descent walking”, International Conference on Medical and [41] P. Peretto, “ An Introduction to the Modeling of Neural Networks”,
Rehabilitation Robotics and Instrumentation, 2014, pp. 336-344. Cambridge University Press, 1992.
[15] T. P. Andriacchi , J. O. Galante, R. W. Fermier, “ The Influence of Total [42] J. D. Bronzino, “ Principles of Electroencephalography”, CRC Press,
Knee-replacement Design on Walking and Stair Climbing”, The Journal Florida, 1995.
of Bone and Joint Surgery, vol. 64, no. 1328-35, 1982. [43] R. Riener, M. Rabuffetti, C. Frigo “Stair ascent and descent at different
[16] G. Bergmann, F.Graichen, A. Rohlmann,“Is Staircase Walking a Risk inclinations”, Gait and posture, 2002, vol. 15, pp. 32-44
for the Fixation of Hip Implants?”, Journal of Biomechanics,Vol. 28, no. [44] E. Isakov, H. Burger, J. Krajnik, M. Grgoric, C. Marincek, ”Influence
5, pp. 535-553. of speed on gait parameters and on symmetry in transtibial amputees”,
[17] R. A. Howard, “Dynamic Programming and Markov Process”, MIT Prosthetics and Orthotics International, 1996, vol. 20, pp 153-158.
Press, 1960. [45] J. A. K. Suykens, J. Vandewalle, “Least squares support vector machine
[18] P. Bremaud, “Markov Chains”, Springer, 1999. classifiers”, Neural processing letters, 1999, vol. 9, no. 3, pp. 293-300.
[19] G. Iyengar, “Robust Dynamic Programming”, Math. Oper. Res., vol. 30, [46] K. VG, “Modeling and analysis of stochastic systems”, Champman and
no. 2, 2005, pp. 257-280. Hall, 1996.
[20] S. Mannor, D. Simester, P. Sun, J. Tsitsiklis, “Bias and Variance [47] K. J. Astrom, B. Wittenmark, “ Adaptive control,2nd eddition”, Addison
Approximation in Value Function Estimates”, Management Science, vol. Wesley, Reading, 1995.
52, no. 2, 2007, pp. 308-322, [48] J. Fox, “Nonparametric Simple Regression: Smoothing Scatterplots”,
[21] M. L. Puterman, “ Markov Decision Processes:Discrete Stochastic Smoothing Scatterplots. Thousand Oaks, CA, 2002.
Dynamic Programming” John Wiley and Sons, Inc, New York, 1994. [49] W. S. Cleveland, S. J. Devlin, “Locally weighted regression: An ap-
[22] J. K. Satia, R. E. Lave, “Markovian Decision Processes with Uncertain proach to regression analysis by local fitting”, Journal of the American
Transition Probabilities”, Operations Research, vol. 21, no. 3, 1973. Statistical Association, 1988,vol. 83, pp.596-610.
[23] C. C. White, H. K. Eldeib, “Markov Decision Processes with Imprecise [50] C. Loader, “Local Regression and Likelihood, 3rd edition”,
Transition Probabilities”, Operations Research, vol. 42, no. 4, 1994. Springer,1999.
[24] N. Hogan, “Impedance Control: An Approach to Manipulation”, Amer- [51] G. Ellis, “Control System Design Guide, 3rd eddidtion”, Elsevier
ican Control Conference, 1984, pp. 304-313. Academic Press, 2004.
[25] L. M. Miller, J. Rosen,“Comparison of Multi-Sensor Admittance Control [52] A. Otto, C. Otto,“ How to design effective priority rules: Example of
in Joint Space and Task Space for a Seven Degree of Freedom Upper simple assembly line balancing”, ELSEVIER, Computers and Industrial
Limb Exoskeleton”, Proceedings of the 3rd IEEE RAS and EMBS Engineering, 2014, vol. 69, pp. 4352.
International Conference on Biomedical Robotics and Biomechatronics, [53] J. Z. Li, H. Gao, “Survey on sensor network research”, Journal of
2010, pp 70-75. Computer Research and Development, 2008, vol. 45, pp. 1-15.
[26] V. Okunev, T. Nierhoff, S. Hirche, “Human-preference-based Control [54] J. L. Elman, “Finding structure in time”, Cognitive Science, 1990, vol.
Design: Adaptive Robot Admittance Control for Physical Human-Robot 14, pp. 179-211.
Interaction”, The 21st IEEE International Symposium on Robot and [55] K. Hornik, M. Stinchcombe, H. White, “Multilayer feedforward net-
Human Interactive Communication, 2012, pp 443-448. works are universal approximators”, Neural Networks, 1989, vol. 2, no.
[27] M. G. Carmichael, D. Liu, “Admittance Control Scheme for Imple- 5, pp. 359366.
menting Model-based Assistance-As-Needed on a Robot”, 35th Annual [56] T. P. Chen, H. Chen, “Universal approximation to nonlinear operators
International Conference of the IEEE EMBS, 2013, pp 870-873. by neural networks with arbitrary activation functions and its application
[28] B. K. Lee, H. D. Lee, J. y. Lee, K. Shin, J. S. Han, C. S. Han, to dynamical systems”, IEEE Transactions on Neural Networks, 1995,
“Development of Dynamic Model-based Controller for Upper Limb vol. 6, no. 4, pp. 911917.
Exoskeleton Robot”, 2012 IEEE International Conference on Robotics [57] J. B. Yang, K. Q. Chen, X. P. Li, “Feature selection for MLP neural
and Automation, 2012, pp 3173-3178. network: the use of random permutation of probabilistic outputs”, IEEE
[29] W. Yu, J. Rosen, X. Li, “PID Admittance Control for an Upper Limb Transactions on Neural Networks, 2009, vol. 20, no. 12, pp. 19111922.
Exoskeleton”, American Control Conference , 2011, pp 1124-1129. [58] A. G. Barto, R. S. Sutton, C. J. C. H. Watkins, “ Learning and
[30] F. Augugliaro, R. DAndrea, “Admittance Control for Physical Human- sequential decision making,” Learning and Computational Neuroscience:
Quadrocopter Interaction”, European Control Conference , 2013, pp Foundations of Adaptive Networks, (1990).
1805-1810. [59] R. Bellman, “Applied Dynamic Programming,” Princeton University
[31] M. Oda, C. Zhu, M. Suzuki, X. Luo, H. Watanabe, Y. Yan, “Admittance Press, Princeton, New Jersey, (1957).
Based Control of Wheelchair Typed Omnidirectional Robot for Walking
Support and Power Assistance”, 19th IEEE International Symposium on
Robot and Human Interactive Communication, 2010, pp 159 - 164.
[32] G. Deodhare, M. Vidyasagar, “ Design of Non-overshooting Feedback
Control Systems”, Proceedings of the 29th Conference on Decision and Abusabah I. A. Ahmed Abusabah I. A. Ahmed is
Control, 1990, pp. 1828-1834. an Associate Professor in College of Engineering,
[33] R. D. Hill, A. C. Eberhard, M. E. Halpern, P. J. Brockwell, “Minimiza- Karary University. He received the M.Sc. degree
tion of peak values of signals in feedback systems”, Proceedings of the and Ph. D. degree in Electronic Engineering from
35th Conference on Decision and Control, 1996, pp. 3212-3217. University of Electronic Science and Technology of
[34] M. Krstic, M. Bement, “Non-overshooting Control of Strict-Feedback China in 2013 and 2017. His research area includes
Nonlinear Systems”, American Control Conference, 2007, pp. 4494-4499. Robotics Control, Predictive and Adaptive Control
[35] A. I. A. Ahmed, H. Cheng, X. Lin, R. Huang, “Motion Planing and Neural Networks. Dr. Abusabah I. A. Ahmed
and Control of Maneuverable Human-Powered Exoskeleton Systems”, has 18 academic publications and he is a reviewer
IEEE/RSJ International Conference on Intelligent Robots and Systems for two journals and many conferences. Now he is
(IROS): Workshop, 2016. teaching Neural Networks and Embedded Systems
[36] K. J. Astrom, R. M. Murray, “Feedback Systems, An Introduction for design for graduate students, and Control Theory, Automatic Control, Modern
Scientists and Engineers”, Princeton University Press, 2008 Control, Optimal Control and Electronic Measurements for junior students at
[37] R. D. Hill, M. E. Halpern, “Minimization of Signed Peak and Absolute Karary University.
Peak Values of Signals in Feedback Systems”, Singapore International
Conference on Intelligent Control and Instrumentation, 1992, pp. 48-53.
 12

Samah H. H. Mohammed Khir Samah H. H. from Karary University in 2020. Her master research
Mohammed Khir is a dean of college of Engineering includes power system stability managing and con-
and Architecture, Shendi University since 2020. She trol approach of doubly Fed Induction Machine for
got her BSc. Degree from Nile Valley University stable operation. Now she is a PhD student at Karary
in Electrical Power Engineering in 2017and worked University. Her current research includes Hybrid Renewable Energy System
as teacher assistant at Shendi University. She got Stability Investigation. Also, the application of Back? to Back Converters in
her master degree in Electrical Engineering, Power both rotary side and Gride Side of Grid Tied Hybrid Renewable power system.