Smart Walker Control through the Inference of the User’s Command

Intentions

M. Martins

, A. Frizera

, C. Santos

and R. Ceres

Universidade do Minho, Gualtar, Braga, Portugal

Departamento de Engenharia Elétrica, Universidade Federal do Espírito Santo, Vitória, ES, Brazil

Grupo de Bioingeniería, Consejo Superior de Investigaciones Científicas,

Crta Campo Real km 0,200, Arganda del Rey, Madrid, Spain

Keywords: Smart Walker, Assistive Mobility, Fuzzy Control.

Abstract: In this work is presented the NeoASAS walker including its conceptual design, implementation and

validation with a new interface approach integrated. This interface is based on a joystick and it is intended

to extract the user’s movement intentions. Eleven healthy users performed preliminary sets of experiments

with the walker, which showed the sensibility of the joystick to extract command intentions from the user.

These signals presented a higher frequency component that was attenuated by a Benedict-Bordner filter.

Then, an approach to the control architecture was developed, in order to obtain stable and safe user

assistance. This control architecture is based on a fuzzy logic control that allows the control of the walkers’

motors. Thus, an assistive device to provide safety and natural manoeuvrability was conceived and offers a

certain degree of intelligence in assistance and decision-making. The motivation is that this will contribute

to improve rehabilitation purposes by promoting ambulatory daily exercises and thus extend users’

independent living.

1 INTRODUCTION

Smart walkers are intended to provide increased

support and assistance during gait. They are adaptive

to their specific application or to the target

population and are designed to continually evaluate

and correct its actions based on its perception of the

users’ needs.

In general, Smart walkers have an integrated

assistive navigation system and sensors to obstacles

detection and there is a concern to allow a stable gait

through different handlebar designs (Frizera-Neto,

2010); (Martins, 2011).

In Smart walkers, the user-walker interface is

intended to interpret the user’s movement intentions

and transform this knowledge into motor commands

(direction and velocity). This research area has

recently witnessed a huge interest in searching for

interfaces that can be intuitive and address the fact

that users are not required being aware of the

intelligent agent behind the driving wheel.

There are many types of interfaces that have

been used in smart walkers. Force sensors are the

most common, as they can be integrated into

handlebars, or in forearm supports (Martins, 2011).

In (Frizera-Neto, 2010) despite good results, users

may present asymmetries during their gait that lead

to different patterns of forces to the same intentions.

These concerns were addressed in (Lee, 2010) by

applying infrared sensors to detect the position of

lower limbs. However, sensors can mix the legs and

therefore make wrong decisions about the users’

intentions.

Despite all the advances in the current state-of-

the-art user-walker interaction field, there are still

many unsolved questions and key areas in

determining user-friendly and efficient interfaces.

Further, it is very important to remember that these

interfaces should not increase the cognitive burden

or cause confusion to the lower limb disable users,

and should be economic.

Additionally, recent studies on the walker

interfaces (Martins, 2011) have not focused on the

characterization of the signals gathered by the

interface sensors, and it is currently lacking an

exhaustive analysis of the main parameters involved

in the interface signal. It is required to identify these

parameters and their connection to the subsequent

458

Martins M., Frizera A., Santos C. and Ceres R..

Smart Walker Control through the Inference of the User’s Command Intentions.

DOI: 10.5220/0004028704580463

In Proceedings of the 9th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2012), pages 458-463

ISBN: 978-989-8565-21-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

algorithms used for detection, recognition and

estimation of user’s commands.

In this work, it is presented a new interface

approach designed to be intuitive and meet usability

aspects. The interface integrates a joystick into the

walker upper base support. Preliminary studies were

conducted with healthy volunteers and no

motorization in the device. An analysis of the

joystick data was performed and user’s navigation

commands were identified. These commands are

going to be used in the guidance of the walker and

recurring to a fuzzy logic strategy, which is

fundamental for an efficient control of the device

during assistive gait. Then, a validation with the

motors on was performed.

This paper is organized as follows. Section II

describes the NeoASAS interface constituted by a

joystick. Section II also presents and discusses the

interaction components acquired with the joystick.

Section III discusses the processing strategy to

extract the signal components related to the user’s

navigation commands. Section IV presents and

discusses the developed control strategy based on a

fuzzy logic system and the achieved results are

present in Section V. Finally, conclusions

are

discussed in Section VI.

2 NeoASAS INTERFACE

The NeoASAS Smart Walker is presented in figure

1. This new robotic walker was built through the

mechanical modification of a conventional four-

wheeled walker. An additional structure was

implemented to integrate the motors and sensors of

the robotic walker, as well as forearm supports.

To program all the implemented strategies on the

walker, it was used the Matlab and PC/104 platform.

2.1 Specifications of the Novel

Interface – Joystick

In this work, the interface consists on placing, at the

centre of the upper base support, a joystick

associated to a spring that is moved according to

user’s manipulation (Figure 1). When the user

begins his gait, he has to slightly move (less than 1

degree) the handlebar through the handles, moving

the joystick, informing the walker which direction

and velocity he wants to take. Hence, the user’s little

efforts are successfully converted into small

movements through this new interface.

To extract and study the signals from the

joystick, it was performed an user study with 11

healthy volunteers, with no history of any

dysfunction on either upper or lower limbs. These

volunteers had to perform simple tasks like moving

forward and then turn left or right. It is noteworthy

that these tests were performed without any

motorized system.

Figure 1: NeoASAS walker and a schematic of the upper

base with the joystick.

The joystick outputs three different signals

(X,Y,Z), measured in Volts that specify the imposed

movement described on the XYZ-axis attached to

the joystick. In this work it is just used Y and Z

signals.

2.2 Interaction Components

Three types of experiments were performed by the

11 healthy users: walking forward, turn right and

turn left (Figure 2). During these two types of

signals were acquired and evaluated – forward (Y)

and rotation (Z). The Y-signal, gives an indication of

the user intention to move forward and according to

the applied force on the X-axis, the signal will have

more or less amplitude, depending on the user’s

command intention to go forward with more or less

velocity. The Z-signal, gives an indication of the

user intention to perform a curve and the signal will

present high or low amplitude depending if the

performed curve is more or less accentuated. The

intention to turn right or left is detected by the sign

of the signal, i.e. turn left causes negative signal and

turn right causes positive signal.

Figure 3 show typical Y and Z joystick data.

Initially, the user is stopped (S1) and both signals

are zero. When the user begins to walk forward (S2),

he pushes forward the handles of the walker and the

Y-signal becomes negative, because the joystick is

moved around the X-axis. The Z-signal continues to

be zero, since the joystick is not rotated. Sometimes,

Z-signal can present some small variations while the

user is moving forward. This is associated to users

that may present more strength in one arm than in

SmartWalkerControlthroughtheInferenceoftheUser'sCommandIntentions

459

the other.When the user turns right (S3), the joystick

is made to move around the Z-axis to the right side

also, presenting a negative Z-signal. At the same

time, in the Y-axis the joystick tends to go to rest,

presenting a zero Y-signal. The next step is to move

forward (S4), and the Z-signal returns to zero while

the Y-signal becomes negative. At the end of the

trajectory, the user stops (S5), and the Y-signal

returns to zero. Z-signal remains in zero.

In the case that the user wants to turn right, the

only difference is that the Z-signal becomes positive

instead negative.

a) b) c)

Figure 2: a) Walking forward, b) turning right and c)

turning left.

Figure 3: Typical raw Y and Z joystick data in the

NeoASAS walker when the user is performing the

following trajectory: S1- The user is stopped; S2- User

starts walking forward; S3- User turns right; S4- User

walks forward; S5 – User stops.

Now, one can conclude that the joystick system

read correctly the user’s command intentions.

However, by observing the characteristics of the

signals Z and Y, it can be identified two main

components of the signals. One component (i)

represents the highest frequency noise caused by the

vibrations of the structure. This component must be

eliminated in real-time. For that, it will be used a

filter which choice will be presented in detail on the

next section. The other component (ii) contains the

information of the walking movement intentions of

the user to guide the walker. This signal will enable

the development of robust and secure control

strategies.

3 FILTERING STRATEGY

The filtering strategy aims to eliminate in real-time

the component (i).

The data that was collected yields that the user’s

commands intentions occur in a frequency range

between 0 and 2 Hz in both Y and Z-signals, and the

higher-frequency components are related to noise. In

figure 4, one can see the spectrum for typical Y and

Z-signals. The Z-signal has more accentuated

higher-frequencies than the Y-signal.

The higher frequency components present in the

signals can be eliminated with forth and back

recursive digital filters, such as Butterworth filters,

without causing phase distortion. However, this

approach is not real-time implementable. As this

technique is not suitable for real-time applications,

this filter will be set as a basis to evaluate the

performance of the chosen filter strategy.

Besides this, the user should not perceive the

delay between his commands and the movement of

the walker. The human perception threshold in

applications like this is known to be around the 200

ms.

In the literature, two types of filters were

identified: g-h filter (Benedict-Bordner and

Critically Damped) and the Kalman filter, and they

are usually called as tracking filter (E.Brookner,

1998).

Figure 4: Example of a Frequency Spectrum for (left) raw

Y-signal and (right) raw Z-signal.

3.1 User’s Movement Intentions

Tracking Filters

g-h filter In this filter measurements are used to

correct the predictions that are made for the signal,

minimizing the estimation error. Formulation is

presented in (Brookner, 1998). This filter presents

two parameters (g,h) that need to be offline tuned.

To this, the Benedict-Bordner Filter (BBF), equation

ICINCO2012-9thInternationalConferenceonInformaticsinControl,AutomationandRobotics

460

(1), and the Critically Damped Filter (CDF),

equation (2) will be applied to select the filter

parameters.

h =

2 − g

(1)

This equation relates g and h, such that the BBF has

one degree of freedom. In g-h filters increasing the

value of g diminishes the transient error. Thus, a

larger g makes the BBF to track higher frequencies.

The (CDF) minimizes the least squares fitting

line of previous measurements, giving old data

lesser significance when forming the total error sum.

This is achieved with a weight factor θ. Parameters

in the g–h filter are related by:

)1(

−=

(2)

Kalman Filter here depicted is the conventional

Kalman filter and is only suitable for linear systems

(E.Brookner., 1998).

Therefore, state vector x(t) is composed by the

variable to be estimated, and its derivative. In the

current problem, is considered the system dynamic

equations:

kkk

xTx

(3)

In these equations it is presented a stochastic model

that considers a first derivative influenced by a

random noise

, i.e. first derivative is not constant.

The equation that links the actual state

and

the measured y

is called the observation equation:

kkk

vxy

(4)

Where v

is the process noise.

The Kalman filter parameters will be the

measurement noise covariance R and the process

noise covariance Q.

In the implementation of the filter, the

measurement noise covariance R is measured prior

to operation of the filter. Measuring the

measurement error covariance R is practical because

generally it is simple to take some off-line sample

measurements in order to determine the average

variance of the measurement noise:

[

]

Its

value is 8.82x10

-5

rad

-2

for the Y-signal noise and

1.3x10

-5

rad

-2

for the Z-signal noise.

The selection of the process noise covariance Q

is formulated based on the first derivative noise,

which affects the estimation of the user’s command

intentions. The value of Q is related to the process

error of the system. Thus, a good choice of Q helps

the filter to estimate more precisely the true state.

It is calculated using of-line measures of the

signal. For each measure the covariance of the signal

is calculated. Finally, the process noise covariance is

the average of all the calculated covariance.

3.2 Evaluation of User’s Movement

Intentions Trackers Filters

The selection of the BBF parameter g, CDF

parameter θ and Kalman filter parameter Q, is

presented in this section.

For this selection the Kinematic Estimation Error

(KTE) was used. KTE evaluates the smoothness,

response time, and execution time of a tracking

algorithm (E.Rocón, 2010) and is expressed by:

σε

+=KTE

(5)

and σ

are the mean and variance of the

absolute estimation error between a desired signal

and the measured signal. The desired signal is

obtained by filtering offline the signals’

measurements with a Butterworth filter.

To select the filters parameters (g, θ and Q), 11

individuals drove the walker without any

motorization executing three different trajectories

with five repetitions each. During these experiments

the signals of the joystick were acquired.

These signals were then introduced off-line in

the 3 filters algorithm using a broad range of g, θ

and Q parameters. The result was processed by the

KTE. The best solutions for each filter, i.e. the ones

with the lowest KTE, were chosen for each user,

experiment and repetition. With these results, it was

calculated the mean of the best 165 solutions for

each parameter, as well as the mean of the delay

between the input and the output for each case.

Table 1 and 2 present the mean values of the best

solutions of g, θ and Q parameters, delay between

the original joystick signal (Y and Z) and the filtered

one and KTE for each joystick signal (Y and Z).

Table 1: Filter Parameters based on the KTE and delay for

the Y-signal. Table provides for mean±standard deviation.

Value KTE (x10

-3

rad/s) Delay (ms)

g 44.20±4.97(x10

-3

) 6.46±0.91 0.5±0.25

θ 0.974±3.85x10

-3

6.81±0.75 1.7±0.96

Q 3.21±0.55(.10

-7

) 9.66±0.86 17.2±1.86

SmartWalkerControlthroughtheInferenceoftheUser'sCommandIntentions

461

Table 2: Filter Parameters based on the KTE and delay for

the Z-signal. Table provides for mean±standard deviation.

Value

KTE (x10

-3

rad/s)

Delay

(ms)

g 16.87±2.51(x10

-3

) 2.93±1.99 23.8±1.70

θ 0.990±1.1x10

-3

2.99±0.11 25.2±1.59

Q 3.26x10

-9

±8.78(.10

-9

) 3.12±0.26 36.2±4.00

As it can be seen in Table 1 and 2, g of the Z-

signal compared with the g parameter of the Y-

signal shows a lower value. Similarly, the average θ

and Q parameters of the Z-signal compared with the

average θ and Q parameters of the Y-signal shows a

higher value. These results were as expected, since

Z-signal required being further filtered.

All filters are of high quality for a human-

machine interaction because the introduced delay is

much more inferior to human perception (200 ms),

not causing prejudice to the human-machine

interaction.

KTE is very low for all filters, being the lowest

one, the BBF’s KTE value, as well as its dispersion.

Additionally, the BBF detains the lowest signals’

delay.

Since BBF presents the lowest KTE for both

signals, one can conclude that it is the best option to

choose for this application.

This can also be seen in an example of joystick

signal in figure 5, where is presented the differences

between BBF and CDF; and figure 6 presents the

differences between BBF and Kalman, as well as the

reference. The BBF shows a higher attenuation on

the oscillations than the CDF and Kalman filters.

Thus, a Benedict-Bordner g-h filter was applied

to the joystick data. The g parameter was chosen to

be 44.29x10

-3

for the Y-signal and 16.87x10

-3

for the

Z-signal. Thus filter has a low computational cost

algorithm, making it a good option to this

application, since it can run in a low cost hardware

with enough robustness for a commercial device.

4 CONTROL STRATEGY

In this section, it is addressed a control strategy

based on fuzzy logic to classify the signals sent by

the joystick and transform them into motor outputs

(direction and velocity), in such way that the walker

drives the motors according to the user’s commands.

The two fuzzy logic inputs will be the Y and Z-

signal.

Figure 5: (left) The superposition of the raw Z with the

results of BBF, CDF and Butterworth; (Right) The

superposition of the raw Y with the results of BBF, CDF

and Butterworth.

Figure 6: (Left) superposition of the raw Z with the results

of BBF, Kalman and Butterworth; (Right) superposition of

the raw Y with the results of BBF, Kalman and

Butterworth.

It was defined a set of membership functions

(MF) for each joystick signal and they were

constituted by Gaussian and bell functions. The variables,

which form the set of MF for the Z-signal and that will

interpret this signal, are divided onto: much left (ME),

little left (LE), zero (Zi), little right (LR) and much right

(VR). Similarly, the variables, which form the set of MF

for the Y-signal and that will interpret this signal, are

divided into: negative (Neg), zero (Ze), little positive (PP)

and very positive (MP). For the motors right (MR) and left

(ML), the output MF set is divided onto: zero (Z), slow (S)

and fast (F).The decision-making rules are presented in

Table 3.

A series of experiments with motorization were

conducted to assess the functioning of the fuzzy

system and allow the tuning of the parameters of the

implemented system.

Table 3: Decision-making rules. Black (white) columns

are related to the left (right) motor.

ICINCO2012-9thInternationalConferenceonInformaticsinControl,AutomationandRobotics

462

5 VALIDATION OF THE

PURPOSED ARCHITECTURE

WITH HEALTHY USERS.

In figure 7, an example of results is shown. The

signals were acquired while a user was performing

the following trajectory: Start to walk, walk forward,

turn left, walk forward and stop. Figure 7 a) shows

that with the addition of the motors in the movement

of the walker, the Y and Z-signals present a more

accentuated noise, comparing with the one saw in

figure 3. However, the results from the filter BBF

are very satisfactory in attenuating the noise

components.

In figure 7 b) it is shown the result obtained

before the BBF filter, as well as an adjustment on

the gains of the signals. The Y and Z-signal were

inverted, amplified and are in the range of [-1,1].

In figure 7 c) the outputs of the fuzzy control

system were smoothed and converted to the range of

[2.5,5] in order to be sent to the low-level control

hardware to command the DC motors.

Figure 7: Results from the system architecture of the

NeoASAS walker. a) raw acquired joystick signals and the

result from the filtering with the BBF filter; b) signals

before and after the amplification and restrictions; c)

output of the fuzzy system (left) and their integration to

then be sent to the control board hardware (right).

Despite the variations of the Y and Z-signal, the

motors present a constant and safe movement.

Therefore, the system is perfectly adjusted to read

the user’s command intentions.

Thus, it was successfully generated a control

strategy which has low computational cost, allowing

a smooth and enjoyable driving, fast response of the

walker and no sense of delay.

6 CONCLUSIONS

In this work it was presented a method of user-

walker interaction to extract the users’ command

intentions. A series of experiments using with

healthy users were performed which showed the

sensibility of the joystick to extract navigation

commands from the user. The proposed control

strategy showed very good results, allowing a

smooth and enjoyable driving, fast response of the

walker and no sense of delay.

ACKNOWLEDGEMENTS

This work is financiered by FEDER Funds and

through Programa Operacional Fatores de

Competitividade – COMPETE and by National

Funds through FCT - Fundação para a Ciência e

Tecnologia under the Project: FCOMP-01-0124-

FEDER-022674.

Work supported by Portuguese Science

Foundation (grant SFRH/BD/76097/2011).

REFERENCES

Martins, M., Frizera, A., Santos, C., Ceres, R., (2011).

Assistive Mobility Devices focusing on Smart

Walkers: Classification and Review. Robotics and

Autonomous Systems. DOI: 10.1016/j.robot.2011.11.

015.

Frizera, A., Gallego, J., Rocon, E., Pons, J., Ceres, R.,

(2010). Extraction of userʼs navigation commands

from upper body force interaction in walker assisted

gait. BioMedical Engineering Online, 9(37), 1-16.

Lee, G., Ohnuma, T., Chong, N., (2010). Design and

control of JAIST active robotic walker. Intel Serv

Robotics.

Brookner, E., (1998). Tracking and Kalman Filtering

Made Easy. John Wiley & Sons, Inc.

Rocón, E., Ruiz, J., Moreno, J., Pons, J., Miranda, A.,

(2010). Tremor Characterization: Algorithms for the

study of tremor time series. Sensors.

SmartWalkerControlthroughtheInferenceoftheUser'sCommandIntentions

463