SmartLeg: An Intelligent Active Robotic

Prosthesis for Lower-Limb Amputees

Remzo Dedić Haris Dindo
University of Mostar University of Palermo, RoboticsLab, DICGIM
Matice hrvatske bb Viale delle Scienze, Ed. 6
Mostar, Bosnia and Herzegovina 90128 Palermo, Italy
remzo.dedic@sve-mo.ba dindo@dinfo.unipa.it

Abstract— In recent years, there has been a worldwide interest in more metabolic energy during level walking compared to the
improvement of mobility of people with lower limb amputation. normal people gait. The additional power needed for the
In spite of significant development of new technologies during the previously mentioned activities must be achieved by means of
last decade, commercial below-knee and above-knee prostheses external energy sources that should be integral prosthetic
are still energetically passive devices. However, many locomotive components.
functions, like walking up stairs and slopes, need significant
power in knee and ankle joints. The additional power for doing Because of the presumption that robust external power
previously mentioned activities needs to be achieved by means of sources are needed, which would be unacceptable for users,
external energy sources, which should be integral prosthetic current lower limb prostheses incorporate knee and ankle joints
components. with different types of passive damping throughout the gait
cycle. There are computer-controlled prostheses nowadays
This paper presents preliminary investigations towards an active which adjust the movement of the prosthesis to any speed and
robotic prosthesis that could potentially enable people with an remarkably decrease step deviations and amputee effort. The
above- or below-knee amputation to perform different types of most modern generation of controlled prostheses with
motions that require power in lower limb joints. Our initial computer-controlled phases of standing and swinging are
prototype, SmartLeg, integrates advanced prosthetic and robotic considered to be an important step forward, since they offer the
technology with the state-of-the-art machine learning algorithms
closest proximity to natural walking. Users of these prostheses
capable of adapting the working of the prosthesis to the optimal
gait and power consumption patterns, and which therefore
are able to walk freely and safely at different speeds, even
provide means to customize the device to a particular user. when walking on rugged ground or downstairs, still they
cannot climb stairs as non-amputees do. The reason for this
Keywords: Active robotic prosthesis, assistive robotics, machine problem lies in the absence of external source of energy, which
learning would have power generation capabilities comparable to an
actual limb and would provide the energy required to lift the
body.
I. INTRODUCTION
Recently there has been a significant interest, both in The goal of the SmartLeg project is to prove a real
academia and industry, in devising advanced solutions for the possibility that a person with an above or below knee
improvement of mobility of people with lower limb amputation can perform different types of motions which
amputation. The main causes of amputation are diseases, require power in knee and ankle joints, like climbing stairs,
trauma and congenital deformities. The most common diseases with hydraulic units whose weight and overall dimensions
that contribute to amputation are vascular diseases, diabetes would not interfere with the comfort of the prosthesis user.
and tumors, usually occurring after age 60. Traumatic SmartLeg prosthesis will have the capability of active
amputations occur in younger and more active population, movement control in knee and ankle joints through the
usually as a result of industrial and motor vehicle accidents. In embedded actuators, which will enable achieving of necessary
countries that are currently at war, or have recent history of war flexion and extension, as well as driving of below knee and
(like Bosnia and Herzegovina, Croatia, Kosovo), besides the above knee components. In addition, Artificial Intelligence
already mentioned causes there is a large population of military based solutions, dealing with learning and adaptation, are being
and civil war victims. Also, there is a risk of injury by investigated in order to adapt to a particular user and to
landmines, which affects mainly civilians and is present long reproduce complex tasks related to human activities. Finally,
after conflicts are over. our aim is to lower the cost of the final product making it
accessible to a large population of users.
In spite of significant development of new technologies
during the last decade, commercial below knee and above knee II. RELATED WORKS
prostheses are still energetically passive devices. However,
many locomotive functions, like walking up stairs and slopes, Scientific and technical innovations have made a significant
need significant power in knee and ankle joints. In addition, it progress towards developing more comfortable, efficient and
has been shown that below- and above-knee amputees use lifelike artificial limbs. Future progress is likely to depend on

978-1-4577-0746-9/11/$26.00 ©2011 IEEE

the interaction between three powerful forces: amputee’s and in a more energy efficient manner [8]. The Otto Bock C-
demands, advances in surgery and engineering, and healthcare Leg takes this a stage further, offering not only symmetry in
funding sufficient to sustain development and application of the swing phase but also an improved security in the stance
technological solutions. Prosthetic technology has advanced to phase. Sensors in the ankle and shin of the prosthesis
a remarkable degree in the past two decades, driven largely by continually assess the position of the leg in space as the
amputees demand. Today, otherwise healthy individuals with amputee is walking. The data are fed into two microprocessors
mid-calf amputation should be able to participate in a full range inside the knee, and the resistance from a hydraulic damper is
of normal responsibilities, to walk without any perceptible adjusted up to 50 times a second, optimizing both the stiffness
limp, and to engage in recreational and sports activities [1]. throughout the entire gait cycle [9], as well as hydraulic
damping, making both walking and stair descent easier [10].
In developed countries the main cause of lower limb However, without the usage of an outside power source, it is
amputation is circulatory dysfunction. The prime reason for
impossible for the body to perform a number of everyday
this is atherosclerosis; although up to a third of patients have activities which involve climbing significant slopes. This
concomitant diabetes. These people are usually in their sixth
makes radically different solutions necessary [11].
decade (or older), and most have additional health problems
that limit their walking ability [2]. This is in sharp contrast with
developing countries, where most amputations are caused by III. SMARTLEG OVERVIEW
trauma related either to conflict or to industrial or traffic The basic idea of our approach is to incorporate additional
injuries [3]. The devastation caused by land mines continues, actuators into an existing, commercially available, passive
particularly when displaced civilian return to mined areas and prosthesis (Fig.1). Such a modified prosthesis will have the
resume agricultural activities [4]. Global extrapolations are capability of active movement control in knee and ankle joints
problematic, but a recent US study states that the amputation through the embedded actuators, which will enable achieving
rate among combatants in recent US military conflict remains of necessary flexion and extension, as well as driving of below
at 14-19 % [5]. knee and above knee components.
The single most critical aspect of any prosthesis is the
quality of the interface between the limb remnants (stump) and
the artificial prosthesis. The portion of the prosthesis that fits
snugly over the limb remnant, the ”socket” determines the
amputee’s comfort and ability to control the artificial limb.
Since the 1980s prosthetic clinicians and researches worldwide
have made breakthroughs in design and materials that have
greatly improved the connection between the socket and stump.
Currently, silicone elastomers are widely used to create a soft
and slightly elastic inner liner, providing a thin, comfortable,
and compliant barrier between the amputee’s skin and the more
rigid, weight bearing portions of the prosthetic socket [6].
Recently researchers have developed a variety of thicker
gel materials that add a measure of cushioning and pressure
dissipation while retaining the benefits of the original liners.
Carbon fiber composites, developed by the aerospace industry, Figure 1: Modified above-knee prosthesis model
are increasingly being used in the artificial limbs, largely
because of their superior strength to weight characteristics. It is intended that embedded actuators, by using the power
There are currently a large number of different design solutions from an outside source, support a disabled person during stair
of the above–knee prosthesis, by which searches are trying to climbing. Indeed, the biggest problem for above–knee
enable and make an easier walking for limb amputees. It is amputees is lack of their own knee, including the lack of
attempted to make above-knee prosthesis movable end enable muscles that enable climbing. The forces normally generated in
its easier use. Large attention is directed towards finding the the knee during climbing are among the largest exerted in the
best construction of mechanical joint of knee and ankle. Helped human body, and are difficult to generate in existing prosthetic
by suitable medical therapeutics and exercises, invalid legs. For that reason, an outside power source is necessary.
individuals “learn” how to walk using their prosthesis. An In addition, active prostheses can provide users with an
important example is given by the so-called “sports prosthesis” increased degree of comfort compared to traditional passive
which gives maximum aid to the sport amputees in achieving actuators by autonomously adapting to their needs. This is
first–class results [7]. achieved by employing advanced learning algorithms from the
The first artificial knee with an “on-board” computer to realm of artificial intelligence and machine learning that, after
improve the symmetry of amputee’s gait across a wide range of a short training period, adjust the on-board control algorithm to
walking speed was developed by Blatchford in the early 1990s. match either observed users’ walking patterns (if the training
Studies have confirmed that these “intelligent prosthesis” offer set can be gathered by sampling data from the other limb), or
amputees a more reliable gait pattern during the swing phase of standard motion patterns from the healthy subjects otherwise.
the gait cycle, permitting them to walk with more confidence By equipping the artificial leg with standard sensors found on
commercial robotic platforms, the onboard microcontroller will
be able to recognize different terrains (their type and slope) and hydraulic system. Fig.2 shows a schematic of the hydraulic
to instantaneously adapt the gait to fulfill users’ expectations. system installed on the existing SmartLeg prototype.
In this paper we focus on the following problems in the
development of a smart active robotic prosthesis (for the B. Control System
description of the mechanical development of the prosthesis The hydraulic system must include an appropriate control
please refer to [12]): system and regulative elements for directing fluid as well as
distributing it towards hydraulic actuators. In the broadest
• development of the optimal hydraulic installation; terms, controls or control systems relate to any device,
• development of the associated control algorithms; component, or combination of components, the resulting output
of which responds to given command input signals.
• development of artificial intelligence solutions dealing Requirements for the control system are:
with intelligent behavior, learning and adaptation able
to reproduce complex walking patterns under different • to collect data of the flexion and extension angles at
conditions. the knee and ankle;
• to identify the force/pressure during the contact the
A. Hydraulic System prosthesis has with the ground;
As said in the introduction, while climbing up (or climbing • to predict the leg motion/gait during the stairs
down) the stairs, the biggest problem for above-knee amputees
climbing based on the sensory input.
is the missing knee power of the existing prostheses. For
example, in the existing electromyography prostheses the Control systems may be with, or without feedback signals.
myoelectric signal is not capable of providing enough power The feedback control system compares the reference set-point
for the above-knee prosthesis, which is the main reason why an to the transduced output and/or controlled variables. The
outside power source is needed. resulting difference between these two signals, or error signal,
is used to reduce the difference between the output and input
This work has the intent of defining a hydraulic system and
toward zero. This error signal can be amplified by means of the
an appropriate control system with an independent source of system components so that the system output almost exactly
power supply. The hydraulics parameters of the linear actuators
follows or matches the system input. Amplification of the
are based on the achieved experimental results when a load was control signal is usually required to actuate solenoid operated
attached to the knee joint as well as on the calculated
hydraulic valves. Electronic controllers that are designed for
theoretical parameters [13]. hydraulic applications will normally include the required
One hydraulic actuator has been installed into the amplification functions.
prosthesis’s knee to provide power to lift the weight of the
body; the other hydraulic actuator has been installed into
prosthesis’s ankle. Both actuators are integral part of the same

6
5

4 4

3 3

2
2

1 Legend:
1 - hydraulic power unit
2 – electro-proportional directional control valve
M 3 – hydraulics installation, hoses, and fittings
4 – quick disconnect
5 - knee actuator
6 - ankle actuator

Figure 2: SmartLeg hydraulic system schematic

 Figure 3: Block diagram of the SmartLeg control system with amplifiers and feedback encoders

Spool type proportional valves will normally have a certain This desiderata becomes a necessity in active upper-limb
amount of spool overlap which produces deadband. For robotic prostheses in which myoelectric signals are not
pressure and flow controls this deadband will occur at the start sufficient to provide a necessary reference input to drive the
of spool movement. For directional valves the deadband will system.
occur around the center position. Spool overlap reduces
leakage in the null position and also provides a greater degree In order to achieve this goal, we plan to adopt a number of
advanced artificial intelligence and machine learning
of safety in power failure or emergency stop situations. The
effect of spool overlap requires that a certain minimum signal techniques for adapting the control algorithm to a particular
user’s needs in different circumstances. To this aim, besides the
level has to be present at the solenoid coil before any
noticeable result occurs in the system. sensors actively used in controlling the robotic prosthesis (cf.
Section III-B), we intend to equip the artificial limb with a
When controlling directional valves using a number of low-consuming sensors commonly used in mobile
multifunctional/dual solenoid driver, the dead band elimination and humanoid robotics such as proximity sensors and sonars
function is normally activated using a dip switch. The amount for detecting the presence of stairs and eventual obstacles and
of dead band jump is typically factory-set, and is application- their distance with respect to the user.
specific.
We plan to adopt a hierarchical control structure in which
When a current passes through a solenoid coil, heat is the low-level controller (cf. Sec. III-B) directly governs the
generated. This heat increases the resistance of the coil. For variables of interest, given a reference set-point provided by an
example, a coil may have a resistance of seven ohms at 20°C, adaptive high-level controller trained to different walking
and a resistance of nine ohms at 100°C. This increased patterns for a particular user. Fig. 4 depicts a high-level view of
resistance causes a reduction in coil power, which results in a the SmartLeg learning and control system.
lower valve setting. To compensate for temperature induced
Typical walking patterns in different circumstances will be
changes, some controller / amplifiers have a function known a
“current feedback”. A current feedback resistor is added to the learned by gathering data coming from various sensors, in
particular those from the other limb. In particular, recurrent
circuit in series with the solenoid coil. This allows the solenoid
current to be proportional to the input signal voltage, and neural networks (RNN) seem a promising choice that allows
learning arbitrary sequences of inputs [14]. RNNs have been
independent of the solenoid resistance. Power supply voltage
must be sufficient to overcome the increased resistance. successfully applied in learning motor control patterns that
govern dynamical systems [15]. Similar problems have also
been studied in generating motion patterns for bipedal
C. Learning and Decision-making System humanoid robots [16].
The “smart” adjective of our system refers to the ability of
the robotic prosthesis to automatically adapt to the user’s
current needs and to maximize its comfort in the everyday life.
 )745/F5023(E4G71A2H74( A. Preliminary studies
.5IC/(7G(5/112E4J(:37C/:J(K<( For optimal construction of hydraulic actuator and related
power source, it is necessary to obtain precise measurements of
the moment in the knee joint stairs climbing. For that purpose,
!"#!$%&'&%()*+,-*%%&-( standard anthropomeasures have been used to compute the
.+/0123$4/56718(92:/;<( required torques. To that aim, we have adopted the wire model
!"#"$%&$"'$()%
of the human body (Fig. 5) used to estimate necessary
*+,#-&"./#0% moments at key points during stair climbing.
%*=$%&'&%()*+,-*%%&-(
(

1"/#$"2% 3,,45(67%

>%?+,(
.@A215%/B(C17:5D/:E:<(

Figure 4: Schematic view of the SmartLeg adaptive control system; High-

level controller will be trained to produce particular motor control sequences
corresponding to various walking patterns; Low-level controller is in charge to
convert these to suitable control signals to drive the plant accordingly; The
choice of which motor sequence to activate will depend both on the current
state of the system, as well as of the contextual information (e.g. presence of
stairs or obstacles, type of terrain, etc.)

Users will undergo typical training sessions in which

signals coming from the sensors will form the training set upon
which the weights and connections of the RNN will be built.
Since different motion patterns are adopted in different Figure 5: Wire model of the human body
circumstances (e.g. climbing vs. walking), we plan to train a
number of RNNs – depending on the operational conditions. For the calculation of the knee moment, five subjects from
Particular care will be given to the stability issues of the the men population have been enrolled and their linear
resulting dynamic system and a number of standard safety tests anthropomeasures (segmental mass and dimensions) have been
will be performed with experts in rehabilitation. collected. Torque measurements have been taken with respect
to six types of stairs (separated according to measure
In order to decide which RNN to activate at a given time, height/width of stairs ground used in domestic environments).
we plan to train a classifier able to recognize the current Wire model of the human body was placed into characteristic
condition based on the available sensory inputs. To this aim, a climbing postures, and a plot of swing angles and the required
number of standard machine learning techniques (e.g. Neural torque during stairs climbing has been obtained (Fig. 6).
Networks, Support Vector Machines, Bayesian classifiers) will
be tested [17].

IV. EXPERIMENTAL RESULTS

In this section we intend to provide a proof-of-concept
related to the mechanical and hydraulic design of the SmartLeg
prototype1. Our initial goal was to assess whether the type of
prosthesis we intend to develop is sufficient to support the type
of motions currently out of reach in commercial prostheses,
such as stair climbing and standing up from a chair. In addition,
we aim at building prostheses able to reproduce motion
patterns similar to those observed in healthy subjects.
A number of experimental data, collected with advanced
motion capture techniques, have been used to drive the design
of mechanical and hydraulic parts and to assess the difference
of walking patterns between healthy subjects and amputees
with our prosthesis installed.
Figure 6: Swing angle vs. knee moment for five different male subjects
1
The adaptive control and learning is still under development and no (Group I to V) under six different conditions (A to F)
results could yet be presented.
Results above were used to define the requirements of the a mean trajectory of the healthy subjects during climbing a
hydraulic actuator. Instead of building a new prosthesis from medium-sloped stairs, while Fig. 9(b) shows the mean
scratch, an appropriate hydraulic cylinder has been installed trajectory of the amputee subjects under identical conditions.
into an existing Endolite (www.endolite.co.uk) prosthesis as
shown in Fig. 7.

b) Amputee subjects with the first

a) Healthy subjects
SmartLeg prototype
Figure 9: Comparison of acquired motion trajectories during moderately-
sloped stairs climbing in healthy and amputee subjects with one DOF
prosthesis

Figure 7: Modified above-knee prosthesis with an appropriate hydraulic

As schematically shown in the figure above there is a huge
actuator in the knee deviation between the walking pattern of an amputee and the
one of a healthy person. This can be readily explained with the
Based on the above experimental data, we have chosen lack of the artificial joint in the ankle of the prosthesis,
Parker-Oildyne 108 Series hydraulic power unit designed for implying that the foot-leg articulations behave as a single rigid
intermittent service (www.parker.com). It’s completely self- body.
contained with a DC motor, gear pump, reservoir, internal
check valve and relief valve. Hydraulic fluid with a viscosity C. Modified prosthesis: adding a hydraulic cylinder in the
of 150 to 300 SUS at 38°C (100°F) is acceptable. Based on the ankle
computed hydraulic parameters of pressure and flow we have To alleviate this problem, an artificial joint has been
chosen a mobile battery-powered hydraulic power unit type installed in the ankle of the existing prosthesis together with an
"Walkpac" produced by the company Enerpac. However, both additional hydraulic actuator (Fig. 10). As the result, the above-
components are slightly oversized with respect to the current knee prosthesis has gained an active rotational degree-of-
torque needs of the SmartLeg prototype and we plan to freedom needed to make the stair climbing more confortable.
significantly reengineer the whole hydraulic unit. The resulting prosthesis is shown in Fig. 11.

Figure 8: Compact hydraulic power supply

B. Motion capture analysis: knee hydraulic actuator

In order to compare the kinematics during stair climbing Figure 10: Additional joint in the ankle of the above-knee prosthesis
between healthy subjects and amputees with our prosthesis we
have performed a motion capture analysis between walking D. Motion capture analysis: ankle and knee hydraulic
patterns under identical conditions. The motion capture system actuators
ELITE has been used to acquire kinematic values of the We have once again comapred the kinematics during stair
characteristic points on the right leg/prosthesis of the subjects climbing between healthy subjects and amputees with the
involved in the experiment. Finally, a model of the trajectories modified version of the SmartLeg prosthesis. This time we
has been reconstructed from the full database. Fig. 9(a) depicts
have adopted POLARIS, a more advanced motion capture mechanical design, aiming at maximizing the functionality of
system based on optical passive markers then the one used in the prosthesis form the engineering point of view (power
the preliminary study (cf. Section IV-B). We installed a total of consumption, stability, etc.), with artificial intelligence-based
nine passive markers Northern Digital Inc. (type 1201115) on adaptive control solutions.
the prosthesis. Two sets of experiments have been performed:
Mechanical design, and the subsequent choice of
the first one with the ankle actuator disabled (resembling the
initial situation with only one degree-of-freedom) and the appropriate actuators capable of fulfilling the goal of achieving
a lightweight prosthesis, has followed from extensive
second one with both actuators being active.
experimentation in which typical climbing patterns of healthy
subjects and amputees have been collected and compared
through advanced motion tracking techniques. The result is an
active 2 DOFs prosthesis with embedded hydraulic linear
actuators and microprocessor-based control. The adoption of
machine learning algorithms capable of learning and
reproducing optimal motor patterns will provide an increased
comfort for the end user and permits to customize the device to
its particular needs.

REFERENCES

[1] M-J.Hsu, H. Nielsen, J. Yack and G. Shurr, “Psihological measurement

of walking and running in people with transtibial amputations with 3
different prostheses”, J. Orthop. Sports Phys. Ther., 1999.
[2] L. Marks and J. Michael, “Artificial limbs”, British Medical Journal,
Vol. 323, 2001.
Figure 11: SmartLeg prosthesis prototype with two linear hydraulic actuators [3] C. Rosenfeld, “Landmines - the human cost”, ADF Health, Vol. 1, 2000.
in knee and ankle [4] P. Meade, J. Mirocha, “Civilian landmine injuries in Sri Lanka”, J.
Trauma, 2000.
Fig. 12 depicts the mean trajectory obtained during stair [5] B. Islinger, R. Kulko, A. McHale, “A review of orthopedic injuries in
climbing with and without the ankle actuator. It is evident that the three recent US military conflict”, Mil. Med. 2000.
the availability of the second degree of freedom in the ankle [6] M. Heim, M. Wershavski, T. Zwas, H. Nadvorna, M. Azaria, “Silicone
makes the trajectory more similar to the one observed in the suspension of external prosthetics; a new era in artificial limb usage”, J.
healthy subjects shown (compare with Fig. 9(a)). Bone Joint Surg.,1997.
[7] S. Zahedi, “The results of the field trial of the Endolite Intelligent
Prosthesis”, in proc. XII International Congress of INTERBOR,1993.
[8] G. Buckley, D. Spence, S. Solomonidis, “Energy cost of walking;
comparison of ‘intelligent prosthesis’ with conventional mechanism”,
Arch. Phiys. Med. Rehabil., 1997.
[9] J. Kastner, R. Nimmerwoll, P. Wagner, “What are the benefits of the
Otto Bock C-Leg? A comparative gait analysis of C-Leg, 3R45 and
3R80”, Med. Orthop. Tech., 1990.
[10] H. Dietl, R. Kaitan, R. Pawlik, P. Ferrara, “C-Leg-un nuevo sistema para
la protetizacion de amputados femorales”, Orthopadie-Technik 3, 1998.
[11] D. Zlatnik, “Intelligently controlled above-knee prosthesis”, in proc. IV
International Conf. Motion and Vibration - MOVIC ’98, Zurich, 1998.
[12] R. Dedic, A. Vucina, M. Hudec, “Simulation of climbing up the various
types of stairs with above knee prosthesis”, International Conference
UPS Mostar, Bosnia and Herzegovina, 2001.
[13] A. Vucina, R. Dedic, D. Milcic, “A new Possibility of Climbing Upstairs
Person with Above - Knee Prosthesis”, 7th International Research /
a) One DOF prosthesis b) Two DOFs prosthesis Expert Conference: "Trends in the Development of Machinery and
Associated Technology" (TMT 2003), September, 2003.
Figure 12: Comparison of acquired motion trajectories between simple 1DOF
prosthesis and advanced SmartLeg 2DOFs prototype during stair climbing [14] J.T. Connor, R.D. Martin, L.E. Atlas, "Recurrent neural networks and
robust time series prediction," IEEE Transactions on Neural Networks,
vol.5, no.2, pp.240-254, March 1994.
V. CONCLUSIONS [15] D. Huh and E. Todorov, “Real-time motor control using recurrent neural
networks,” in Proceedings of the 2nd IEEE Symposium on Adaptive
In this paper we have presented preliminary investigations Dynamic Programming and Reinforcement Learning, 2009, pp. 42–49.
towards an active robotic prosthesis that could potentially [16] W. Huang, C.-M. Chew, Y. Zheng, and G.-S. Hong, “Pattern generation
enable people with an above- or below-knee amputation to for bipedal walking on slopes and stairs,” in Proc. of the IEEE-RAS Intl.
perform different types of motions that require power in lower Conf. on Humanoid Robots, 2008, pp. 205 –210.
limb joints such as slope walking or stairs climbing. Our initial [17] C.M. Bishop. Pattern recognition and machine learning. Springer New
prototype, called SmartLeg, aims at combining advanced York, 2006