IEEE/ASME TRANSACTIONS ON MECHATRONICS, 1

Discover2Walk: A Multi-Module Cable-Driven

Robotic Platform for Gait Assistance in Small
Children with Cerebral Palsy
Pablo Romero-Sorozabal, Gabriel Delgado-Oleas, Jorge Muñoz Yañez-Barnuevo Member, IEEE,, Álvaro
Gutiérrez , Senior Member, IEEE, Eduardo Rocon , Member, IEEE

Abstract—Robotic platforms have made significant strides in strategies [11], have shown promising results in adult gait
adult gait rehabilitation, yet similar advancements for young rehabilitation.
children remain limited. To address this gap, we developed the Despite these advancements, equivalent systems for pedi-
Discover2Walk (D2W), a multi-modular robotic system tailored
for early motor rehabilitation in children with Cerebral Palsy. atric patients, particularly toddlers, remain underdeveloped
The D2W integrates cable-driven modules with a treadmill, pro- [12]. Early intervention is crucial for improving motor out-
viding targeted actuation at the pelvis and ankles. Built around comes, yet the rapidly changing physiology of toddlers re-
a bio-inspired hierarchical control architecture and synchronized quires systems that are adaptable and specifically tailored
via Robot Operating System 2, the system is designed to offer to their developmental needs [13]. Pediatric rehabilitation
personalized gait patterns tailored to individual needs. Technical
validation included kinematic tracking and force estimation presents unique opportunities, especially for toddlers with
studies, showing the system’s ability to monitor and control Cerebral Palsy (CP), a leading cause of motor impairment in
movement and force application. To assess the practical use of the children [14].
entire system, an open-loop tracking validation was conducted Nonetheless, some robotic devices have been developed
using a custom-designed 0.9-meter-tall dummy, confirming the for pediatric rehabilitation, such as overground platforms like
platform’s viability in simulated gait assistance scenarios. These
findings suggest that the D2W system could serve as both a CPWalker [15] and Pro-GaitT [16], but these systems often
rehabilitation tool and an observational platform for studying provide limited assistance and have size restrictions that do
pediatric motor development, playing a potential role in early not fully address the needs of small children and toddlers
intervention strategies for children with motor impairments. [12]. Other approaches, such as mobile robots for crouch gait
Index Terms—Rehabilitation robotics, Exoskeletons, Cable- improvement [17], crawling promotion [18], and bodyweight
driven Parallel Robots, Cerebral Palsy support systems for motor learning [19], offer benefits but
often lack the necessary degrees of assistance and adaptability
required for this young population.
I. I NTRODUCTION
To address these challenges, we designed and build the
In recent decades, robotic platforms for gait assistance Discover2Walk (D2W) platform, a fully integrated, multi-
and rehabilitation have advanced significantly, particularly for modular robotic system designed for early motor rehabilitation
adult populations. Systems such as Lokomat [1], LOPES [2], in toddlers. The D2W system features a cable-driven archi-
Gait Trainer [3], Innowalk [4] or ALEX [5] have proven tecture combined with a treadmill, providing actuation at the
effective in restoring or enhancing movement in individuals pelvis and ankles.
with neuromotor gait impairments. This paper presents the design and technical validation
A growing trend in robotic rehabilitation focuses on the of the D2W system. Central to the D2W is a bio-inspired
use of flexible actuators, such as exosuits or end-effector hierarchical architecture [20] that mimics the organization of
cable-driven parallel robots (CDPRs), which address some the Central Nervous System (CNS). This architecture includes
known limitations of rigid exoskeletons, such as inertia, joint three levels: a high-level module for system configuration
misalignments, and difficulties in adapting to different mor- and user interface, a middle-level module with a gait pattern
phologies [6], [7]. Soft wearable cable-based robotic assistance generator to guide system behavior, and low-level modules
platforms [8], CDPR for bodyweight support and sensory controlling the cable-driven components and treadmill. The
feedback [9], or to explore effects in perturbations during integration and synchronization of these modules are managed
bodyweight support [10], or to study force control assistance through ROS2, ensuring coordinated and responsive system
operation.
Pablo Romero-Sorozabal, Gabriel Delgado-Oleas, Jorge Muñoz Yañez-
Barnuevo and Eduardo Rocon are with the Centro de Automática y Robótica, In this paper it is provided an overview of the system’s
Consejo Superior de Investigaciones Cientı́ficas – Universidad Politécnica de mechanical and electronic architecture, and the methodologies
Madrid (CSIC-UPM), 28040 Madrid, Spain. (e-mail: [email protected]). employed for validating its kinematic and force estimation
Gabriel Delgado-Oleas is also with Universidad del Azuay,
Cuenca,Ecuador. capabilities. To assess the system’s performance in gait reha-
Álvaro Gutierrez is with ETSI Telecomunicación, Universidad Politécnica bilitation scenarios for small children, we conducted technical
de Madrid, Madrid, España. experiments using a custom-designed, 14-degree-of-freedom
passive dummy that mimics the size and characteristics of a
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 2

Fig. 1. D2W Robotic Platform. The system comprises two primary cable-driven modules (pelvis and ankles) for gait monitoring and assistance, integrated
with a treadmill. All components are synchronized through a bio-inspired hierarchical control architecture, where a gait pattern generator, adapted to the
patient, coordinates the system’s modules.

0.9-meter-tall child. These tests evaluated the D2W’s ability (PRIM, Spain), which features a U-shaped aluminum frame
to assist motion, demonstrating its potential not only in aiding that serves as the EE of the SCDPR, directly interacting with
rehabilitation but also as a platform for studying pediatric the patient’s pelvis (see Fig. 1 a)). Additionally, at the pelvic
motor development. module it is placed an inertial measurement unit (IMU) that
The results of the validation tests suggest that the D2W includes a triaxial accelerometer, gyroscope, and geomagnetic
system is a promising tool for early intervention in CP sensor (BNO055 IMU, BoschSensortec, Germany) to monitor
and other pediatric motor disorders. While further research the rotations of the pelvis.
and clinical trials are necessary, the D2W platform offers
a foundation for developing more effective and personalized 2) Ankles Modules: At the level of the patient’s lower
rehabilitation strategies tailored to the needs of young children. limbs, two CDPRs are attached to the robotic frame (see
Fig. 1 b)). These CDPRs operate within the sagittal plane
II. S YSTEM D ESIGN of each leg, assisting the patient’s ankles throughout the
The D2W is a multi-module cable-driven robotic platform gait cycle, (see Fig. 2 b)). The assistance is provided by
designed to assist gait of small children with cerebral palsy controlling the cable lengths through an Assistance-As-Needed
providing synchronized assistance at pelvic and ankles joint (AAN) strategy based on the principle of impedance control
levels (see Fig. 1). The system is built around a rigid frame [21]. The motion generated by the CDPRs is transmitted
made of aluminum profiles, which surrounds the patient and to the patient’s ankles using a commercial foot strap (SYL
serves as the mounting structure for the platform’s robotic Fitness®) that is connected to four braided Kevlar® wires with
modules. a 0.5mm diameter (SeaKnight®) (see Fig. 1 d)). These wires
are actuated by four direct-driven brushless DUAL SHAFT
A. System Modules MOTORS – D6374 150KV (OdriveRobotics, United States),
1) Pelvic Module: The pelvic module, is mounted at the top which are positioned at the corners of each CDPR workspace
corners of the rigid frame and functions as a suspended cable- (see Fig. 2 b)´). The motors are controlled using two Odrives
driven parallel robot (SCDPR) designed to assist and monitor v3.6 motor drivers (OdriveRobotics, United States).
the motion of the pelvis, (see Fig. 1 b) and Fig. 2 a)). This
module consists of four XH540-W150-T servomotors (ROBO- 3) Treadmill Module: The treadmill module facilitates user
TIS, United States) which control the length of the cables, (see movement and supports the robotic frame (see Fig. 1 b)). This
Fig. 1 c)). These cables are made from nylon-coated 7x19 316 module consists of a modified treadmill (DOMYOS T540A,
marine-grade stainless steel (TECNI, United Kingdom), and Decathlon, France) designed for external control. The real-
their tension is measured using DYMH-103 load cells (CALT time velocity of the treadmill serves to synchronize the gait
sensor, China). The motion generated by the actuators is patterns during rehabilitation, ensuring coordinated movement
transmitted to the user via a modified commercial hip orthosis across all robotic modules.
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 3

Fig. 3. D2W System Hierarchy and Communication Architecture. This

diagram illustrates the three-tiered control structure and data flow within the
D2W platform. The high level encompasses the clinical interface, allowing
therapists to configure and monitor robotic therapy sessions. The middle level
Fig. 2. Geometric Models and Interaction Forces in the D2W EEs. Top images features a gait pattern generator that dynamically adapts and synchronizes
shows the geometric configurations for the pelvic module a) and ankle module movements across the system. The low level controls the individual modules,
b). The bottom row illustrates the interaction forces observed in the pelvic including pelvic assistance, ankle actuation, and treadmill operation. All com-
orthoses c), in the customized pulley system for tension measurement d), and ponents operate as ROS2 nodes on single-board computers, communicating
in the ankle othosis e). via Ethernet/WiFi.

B. System Architecture Delta 864 single board computers (SBC) with 8GB RAM and
The modular bio-inspired multi-level communication archi- 64GB eMMC (LattePanda, China), running Ubuntu 20.04 with
tecture is based on Robot Operating System 2 (ROS2) to a PREEMPT RT patch to support soft real-time performance,
synchronize its three modules. This architecture is designed to ensuring the system’s responsiveness and reliability.
mimic the organization of the Central Nervous System (CNS)
by dividing the control structure into three hierarchical levels: III. K INEMATIC M ODELING
high, middle, and low [20] (see Fig.3). The aforementioned D2W modules are designed to monitor
The high level serves as the user interface, responsible for and assist the user’s pelvis and ankles during gait. In this
managing and visualizing spatiotemporal gait parameters and section, we detail the kinematic model of each cable-driven
regulating key variables such as the level of gait assistance module, which forms the basis for assisting the pelvic and
and treadmill speed. This level is accessible via a web-based ankles motion during gait (see Fig. 2 a) and b)).
interface, developed with HTML, NodeJS, JavaScript, CSS,
and ROS2, and hosted locally on the robot’s computer.
The middle level functions as the central gait pattern gen- A. Inverse Kinematics
erator (CPG), which adapts the gait trajectories based on the The operation space O, where the user’s joints are assisted,
selected user’s morphological characteristics and the real-time is defined by the Cartesian positions and orientations of these
velocity of the treadmill [22]. These adapted joint positions joints. These positions and orientations coincide with those of
and orientations are published within the ROS2 environment, the robotic end-effectors
 (EEs), which are represented by the
providing reference set-points for the D2W modules to ensure state vector q̄k = x̄k θ̄ k , where k denotes the specific joint,
synchronized assistance during gait. such as the pelvis or an ankle.
The low level handles real-time tracking and control of the The configuration space C, which directly relates to how
gait patterns. It includes the controllers for the pelvic, ankle, the positions and orientations are achieved, is defined by
and treadmill modules, implemented in C++ and Python, the lengths of the n cables connected to each end-effector
which operate as ROS2 nodes. These controllers track the (EE). These cable lengths are represented as the vector ¯lk =
target positions generated by the middle level and adjust the [lk,1 . . . lk,n ]> .
modules based on position and velocity feedback. The relationship between O and C is governed by the
The entire architecture is implemented on LattePanda 3 inverse kinematics The relationship between the operational
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 4

Fig. 4. D2W Motion Tracking Validation. a) Experimental set-up used to compare the D2W’s EE estimated positions to VICON system mesurement. b):
Pelvic position comparison along gait like patterns. c): Ankle position comparison along gait like patterns.

space O and the configuration space C is governed by the where Φ(¯lk,i ) is the total cable lengths estimation error for the
inverse kinematics (IK). The IK involves determining the k joint. The positions of the EE q̄k that minimize this error
cable lengths l̄k that correspond to the desired position and are considered the solution to the FK problem.
orientation of each EE. Calculating these cable lengths is fa- 1) Pelvic Module: The pelvic module presents six degrees
cilitated by the structure matrix A(q̄k ) ,(1), which is derived by of freedom (DOF) and four cables. This makes the FK problem
computing the Jacobian of the standard cable length equation inherently underdetermined when considering only the cable
presented in (2). length equations, as ¯lP ∈ R4 and q̄P ∈ R6 . To solve
  this, we reduce the number of unknowns by incorporating
ūk,1 (b̄k,1 × ūk,1 ) an inertial measurement unit (IMU) to measure the rotational
 ūk,2 (b̄k,2 × ūk,2 ) 
components of q̄P . This approach converts the problem into
A(q̄k ) =  . (1)
 
..
 .. an overdetermined system, allowing for a unique solution to

. 
ūk,n (b̄k,n × ūk,n ) the pelvis position.
2) Ankle Modules: The ankle modules is designed to assist
lk,i = kx̄k + R(θ̄k )b̄k,i − āk,i k (2) within the saggital plane, providing three DOF: translation
along the X-axis and Z-axes, and rotation along the Y-axis
(ankle dorsiflexion-extension). Given this considerations, the
x̄k + R(θ̄k )b̄k,i − āk,i
ūk,i = (3) cable lengths vector ¯lA ∈ R4 and the position vector q̄A ∈ R3 ,
lk,i the system also defined as overdetermined allowing for a
where R(θ̄k ) is the rotation matrix, b̄k,i is the anchor point unique solution.
between the k-th EE and the i-th cable, āk,i is the position of
the i-th winch, origin of the i-th cable and ūk,i represents the
unit vector along the direction of the i-th cable. C. Motion Tracking Validation
This formulation fully defines the IK of each cable-driven To validate the accuracy of the D2W to track the position of
modules in the platform as: its EEs, we used a precomputed gait pattern to move the EEs
¯lk = A(q̄k )q̄k simulating typical pelvic and ankles motion in small children.
(4)
The positions of the pelvic frame (dimentions 20x20 cm) and
ankle frame (12 cm long segment) were continuously recorded
B. Forward Kinematics using a VICON system (Vicon Motion System, UK), which
Determining the position of EE from the cable lengths in comprised 12 cameras capturing data at 100 Hz, and compared
a CDPR is a non-trivial problem since it is defined as the to the positions tracked by the D2W (see Fig. 4). Reflective
intersection of n spheres, where n represents the number of markers placed on the pelvic and ankle frames provided highly
cables [23]. This forward kinematics (FK) problem is typically accurate three-dimensional position data with a resolution of
solved using numerical methods to approximate the solution. 0.01 mm (see Figure 4 a)).
A common approach involves using optimization algorithms to The results of this experiment showed that for the pelvis
minimize ψk,i , the error function representing the difference trajectories, the root mean square error (RMSE) values were
between the measured and computed squared cable lengths, 0.56 mm, 2.75 mm, and 1.72 mm for the X, Y, and Z
[23]. axes, respectively, with corresponding correlation coefficients
of 0.95, 0.99, and 0.99 (see Fig. 4). For the ankle module, the
n
X RMSE values were 16.38 mm, 17.37 mm, and 20.21 mm for
Φ ¯lk,i = min


ψk,i (5) the X, Y and Z respectively, with correlation coefficients of
i=1
0.99 for the X-axis, undefined correlatio for the Y-axis (due
to zero variance in D2W estimation) and 0.98 for the Z-axis,
2
ψk,i ¯lk,i , q̄k = āk,i − x̄k − R(θ̄k )b̄k,i 2


− lk,i =0 (6) respectively, see right image in Fig.4.
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 5

Fig. 5. D2W Force Estimation Validation. a): Experimental set-up using the 3-axis force sensor T501 fixed at the D2W’s EEs. b): Comparison between the
D2W and T501 sensed forces at the pelvis EE. c): D2W generated forces compared to the sensed ones with the T501 force sensor at the ankles EE.

The differences observed between the position estimations B. Force Distribution

of the ankle and pelvis modules stem primarily from the The force distribution problem involves determining the
simplifications made in their respective FK models. The pelvis cable tensions required to generate a desired f¯k at the EE.
is considered as a six-DOF system; however, to address the This process relies on solving (7), which typically involves
underconstrained nature of its FK problem, an IMU is used using minimization algorithms [24]. The goal of this opti-
to simplify the system, resulting in a determined system with mization problem is to find the cables’ modules tension vector
a unique solution. Consequently, the errors in pelvis position T̄ = [T1 . . . Tn ]> that produces the desired f¯k while adhering
estimation are primarily attributed to inaccuracies in the IMU to physical constraints. These include ensuring that the cables
or cable length measurements. do not push (i.e. maintaining a minimum cable tension T̄min )
In contrast, the ankle module, which in reality operates also and remain within the maximum allowable tension T̄max . The
with 6 DOF, is simplified to a 3-DOF system by considering D2W addresses this by employing a bounded least squares
only movement in the sagittal plane. This simplification in- solution defined as:
herently excludes motion in other planes, which affects the
accuracy of the position estimations, as shown in the middle
min kA(q̄k )> T̄k − f¯k k2 (8)
image of Fig. 4 c). Despite these simplifications, the D2W T̄
platform demonstrates reliable tracking of both pelvis and subject to:
ankle positions, with errors ranging from 0.56 mm to 20.21
T̄min ≤ T̄ ≤ T̄max (9)
mm.

IV. F ORCE A NALYSIS C. Force Analysis Validation

The effective operation of the D2W platform relies not only To evaluate the D2W performance in estimating and gen-
on its ability to monitor but also to accurately track and apply erating forces in the EE, experiments were conducted to
forces at the pelvis and ankles during gait. In this section, compare the generated and sensed gait-like forces by the D2W
we outline the methodology used to estimate and apply these with measurements from a 3-axis force sensor with a sensing
forces. capacity of 300N at 80 Hz and a resolution of 0.2N, T501
(Right, China) (see Fig. 5 a)).
A. Force Estimation For the pelvic module, the experiment involved fixing the
To estimate the applied forces at the EE, the D2W re- rigid pelvic frame to the T501 sensor and applying gait-like
constructs them based on the estimated EE position and the forces and partial body weight support by controlling the
sensed cables module tensions. The tension in each cable, motors with periodic currents (see Fig. 5 b)). The generated
Tk,i , is measured using a custom-built pulley mechanism that tensions ranged from -5N to 5N in the X-axis, -14N to 14N in
integrates a load cell to measure the reaction force module, the Y-axis, and from 55N to 85N in the Z-axis. The D2W used
F , produced by the cable tension (see Fig. 2 d)). the sensed pelvic cables tensions, T̄P , and the computed EE
Since this mechanism is fixed, and the cable and pulley positions, q̄P , to obtain f¯P using (7). The D2W estimations
masses are considered negligible compared to the EE mass, were compared to the measurements from the T501 sensor.
the tension module of each cable can be estimated as twice The RMSE values found were 0.57N, 2.86N, and 1.97N for
the vertical reaction force sensed by the load cell (T = 2F1 ), the X, Y, and Z axes, respectively, with a correlation coefficient
as defined in Fig. 2 d). of 0.98 across all axes.
Once is obtained the tension module of each cable, T̄ = Similarly, for the ankle modules the rigid frame was fixed to
[T1 . . . Tn ]> , it is computed the estimated wrench (force and the T501, as shown in the lower image of Fig 5 a). The CDPR
torque), f¯k , using (1). was then used to generate gait-like force assistance trajectories
at the EE, which were compared to the measurements from
f¯k = A(q̄k )> T̄k (7) the T501 sensor (see Fig 5 c)). The forces were generated by
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 6

Fig. 6. Multi-DOF Lower Limb Dummy. The dummy used for validation of the D2W features 3 DOF at the hip, 1 DOF at the knee, and 3 DOF at the ankle
allowing flexion/extension, internal/external rotation, and abduction/adduction movements.

Fig. 7. Open-loop Position Tracking of the D2W. The platform modules adjust cable lengths based on EE positions provided by the gait pattern generator. The
pelvic module uses inverse kinematics (IK) and the embedded PID controllers in the servo motors to control the cable lengths. The ankles module employs
an AAN control strategy to assist ankle movement, utilizing impedance-based control over the computed cable lengths corresponding to the ankle positions
derived from IK.

controlling the CDPR motors in current control mode, were the evaluate the resulting effects on the EE positions, particularly
controlled currents, ūA , were computed based on the known at the pelvis and ankles.
drum radii of the motors, rd , the torque constant, kt , and The experiment involved a full-system validation where a
the estimated cable tensions distribution obtained using (8) defined treadmill speed of 0.5 km/h was imposed and the
as: ūA = rd T̄A kt . The final generated forces ranged from gait trajectories for the pelvis and ankles generated by the
-27N to 42N in the X-axis, constant 0N in the Y-axis and - gait pattern generator [22]. To simulate a gait rehabilitation
21N to 38N in the Z-axis (see Fig. 5 c)). The found RMSE scenario with a small child we developed and used a multi-
values were of 6.07N, 5.37N and 3.7N for the X, Y and Z DOF lower limb dummy designed to replicate human lower
axes, with correlation coefficients of 0.97, undefined and 0.97, limb kinematics, featuring 3 degrees of freedom at the hip, 1
respectively. DOF at the knee, and 3 DOF at the ankle (see Fig. 6). This
The RMSE variations between the ankle and pelvic modules dummy, with a weight of 12 kg and a leg length of 65 cm,
are likely due to the accumulated position errors when using served as a realistic testbed to examine the system’s ability to
(5) to compute q̄k , which in turn affects (8). Additionally, in monitor and assist human gait. The D2W system controlled
the ankle module, some discrepancies also arise from observed the cable lengths connected to the dummy’s pelvis and feet to
forces in the Y-axis, which influence the force estimations in follow the trajectories as illustrated with a circle in Fig. 7.
the other axes, (see middle image of Fig. 5 c)). For the pelvis module, IK calculations provided the desired
cable lengths necessary to achieve specific pelvis movements.
V. F ULL S YSTEM O PEN -L OOP VALIDATION These cable lengths were then regulated by PID controllers
embedded within the pelvic servomotors. The accuracy of the
To validate the performance of the D2W system, we con- system was measured by evaluating the RMSE and correlation
ducted an open-loop position tracking validation focused on between the controlled and reference cable lengths, as shown
assessing its effectiveness in gait assistance scenarios. The in Table I. The results demonstrated that the RMSE values
primary objective was to determine how accurately the system for the four cable lengths ranged between 0.31 cm and 0.33
could follow cable lengths calculated through IK and to cm, with correlation coefficients between 0.91 and 0.92. The
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 7

Fig. 8. Spacial Position Tracking of the D2W. Representation of the open-loop position tracking trial performed with the D2W when using the multi-DOF
lower limb dummy. Left Image: Three-dimensional representation of the pelvis and ankles positions. Middle Image: Detailed view of the three-dimensional
trajectory of the pelvis. Right Figure: Detailed view of the right ankle trajectory over the sagittal plane.

Fig. 9. Time-dependent Position Tracking of the D2W. Time-dependent representation of the open-loop position trial performed with the D2W system using
a multi-DOF lower limb dummy. The left graphs shows the measured position and angle for the pelvis module (Y-axis translation and yaw rotation) alongside
the corresponding controlled cable lengths. The right side illustrates the measured position for the right ankle module (X and Z axes) along with the controlled
cable lengths.

system’s effect on EE positions was also analyzed, particularly TABLE I

focusing on Y-axis translation (pelvic lateral motion) and yaw S UMMARY OF RMSE AND C ORRELATION C OEFFICIENTS FOR P ELVIC
AND A NKLE P OSITIONS IN THE O PERATION AND C ONFIGURATION SPACE
rotation (pelvic rotation). The RMSE for the Y-axis translation
was 0.83 cm, with a correlation of 0.90, while the yaw rotation Module Movement RMSE Correlation
exhibited an RMSE of 1.1 degrees and a correlation of 0.98 Cable Length 1 0.31 cm 0.91
Cable Length 2 0.32 cm 0.92
(see Fig. 8 and 9). C space
Cable Length 3 0.33 cm 0.91
Pelvis
Cable Length 4 0.31 cm 0.92
Similarly, the ankle modules used IK to determine the O space
Y-axis 0.83 cm 0.90
required cable lengths for achieving desired joint movements. Yaw Rotation 1.1 deg 0.98
Cable Length 1 4.17 cm 0.89
An AAN control strategy, incorporating impedance control, Cable Length 2 4.21 cm 0.91
C space
was used to manage the interaction between the cables and Ankles
Cable Length 3 4.99 cm 0.94
the dummy’s joints (see Fig. 7). Since impedance control is Cable Length 4 4.28 cm 0.90
X-axis 5.77 cm 0.87
designed to assist motion rather than strictly enforce trajec- O space
Z-axis 3.42 cm 0.84
tories, and considering the dummy’s complete passivity, the
greater errors observed at the ankles, mean RMSE of 5.77
cm for the X-axis and 3.42 cm for the Z-axis, see Table I, VI. C ONCLUSION
were expected with correlations of 0.87 and 0.84 respectively.
The cable length RMSEs ranged from 4.17 cm to 4.99 cm, This paper presented the D2W system, a multi-modular
with correlations between 0.89 and 0.94. These results indicate robotic platform designed for early motor rehabilitation in
that while the trajectories are similar, there is a delay in the toddlers, integrating cable-driven modules with a treadmill
response, characteristic of the impedance control approach (see system. The platform is built around a bio-inspired hierarchi-
Fig. 9). cal control architecture, synchronized through ROS2, which
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 8

allows for responsive and personalized rehabilitation protocols. [4] Å. B. Tornberg and K. Lauruschkus, “Non-ambulatory children with
Technical validation demonstrated the platform’s ability to cerebral palsy: effects of four months of static and dynamic standing
exercise on passive range of motion and spasticity in the hip,” PeerJ,
track movements and apply forces. The pelvic module was vol. 8, p. e8561, 2020.
tracked with RMSE values between 0.56 mm and 2.75 mm, [5] S. K. Banala, S. H. Kim, S. K. Agrawal, and J. P. Scholz, “Robot assisted
while the ankle modules showed RMSE from 16.38 mm to gait training with active leg exoskeleton (alex),” IEEE Transactions on
Neural Systems and Rehabilitation Engineering, vol. 17, pp. 2–8, 2 2009.
20.21 mm. Force application and sensing validations where [6] C. Siviy et al., “Opportunities and challenges in the development of
also performed observing RMSE between 0.57 N and 2.86 N exoskeletons for locomotor assistance,” Nature Biomedical Engineering
when sensing forces at the pelvis and RMSE values from 3.7 2022 7:4, vol. 7, pp. 456–472, 12 2022.
[7] A. Schiele and F. C. V. D. Helm, “Kinematic design to improve
N to 6.07 N when generated with the ankles module. ergonomics in human machine interaction,” IEEE transactions on neural
In the full system validation, the D2W platform’s ability systems and rehabilitation engineering : a publication of the IEEE
to follow IK calculated cable lengths was assessed through a Engineering in Medicine and Biology Society, vol. 14, pp. 456–469,
12 2006.
open-loop strategy. The pelvis module exhibited RMSE values [8] L. N. Awad et al., “A soft robotic exosuit improves walking in patients
between 0.31 cm and 0.33 cm for cable lengths, with corre- after stroke,” Science Translational Medicine, vol. 9, 7 2017.
sponding RMSE values of 0.83 cm for Y-axis translation and [9] M. Bannwart, M. Bolliger, P. Lutz, M. Gantner, and G. Rauter, “System-
atic analysis of transparency in the gait rehabilitation device the float,”
1.1 degrees for yaw rotation. The ankle modules, utilizing an 2016 14th International Conference on Control, Automation, Robotics
AAN strategy, with RMSEs of 5.77 cm for the X-axis and 3.42 and Vision, ICARCV 2016, 2016.
cm for the Z-axis. These findings underscore the platform’s [10] V. Vashista, M. Khan, and S. K. Agrawal, “A novel approach to apply
gait synchronized external forces on the pelvis using a-tpad to reduce
potential for improving pediatric motor rehabilitation with walking effort,” IEEE Robotics and Automation Letters, vol. 1, pp. 1118–
further refinements. 1124, 7 2016.
The results suggest the potential viability of the D2W [11] J. Fang, M. Haldimann, L. Marchal-Crespo, and K. J. Hunt, “Devel-
opment of an active cable-driven, force-controlled robotic system for
platform for gait assistance in small children, demonstrating walking rehabilitation,” Frontiers in Neurorobotics, vol. 15, p. 49, 5
that a cable-driven robotic platform could effectively support 2021.
pediatric gait rehabilitation. [12] S. Lefmann, R. Russo, and S. Hillier, “The effectiveness of robotic-
assisted gait training for paediatric gait disorders: systematic review,”
Future efforts will focus on continue enhancing the control Journal of NeuroEngineering and Rehabilitation, vol. 14, no. 1, p. 1,
strategies by closing the control loops using positions com- 2017.
puted through FK solutions and controlling them by applying [13] I. Novak et al., “Early, accurate diagnosis and early intervention in
cerebral palsy: Advances in diagnosis and treatment,” JAMA Pediatrics,
the studied tension distribution methods and system dynamics vol. 171, no. 9, pp. 897–907, Sep. 2017, erratum in: JAMA Pediatr.
[25]. Moving from technical validation to clinical trials will be 2017 Sep 1;171(9):919. doi: 10.1001/jamapediatrics.2017.3169.
an essential step in fully establishing the platform’s efficacy in [14] C. Morgan et al., “Effectiveness of motor interventions in infants with
cerebral palsy: a systematic review,” Developmental Medicine & Child
real-world therapeutic settings. Overall, this study highlights Neurology, vol. 58, no. 9, pp. 900–909, 2016.
the potential of D2W as a promising solution for pediatric gait [15] L. F. Aycardi et al., “Evaluation of biomechanical gait parameters of
rehabilitation, while also recognizing the need for continued patients with cerebral palsy at three different levels of gait assistance
using the cpwalker,” Journal of neuroengineering and rehabilitation,
research and development. vol. 16, 1 2019.
[16] A. J. McDaid, “Design, analysis, and multicriteria optimization of an
ACKNOWLEDGMENTS overground pediatric robotic gait trainer,” IEEE/ASME Transactions on
Mechatronics, vol. 22, pp. 1674–1684, 8 2017.
This work was carried out within the framework of [17] E. J. Park et al., “Design and preliminary evaluation of a multi-robotic
the Discover2Walk project “Development of a robotic plat- system with pelvic and hip assistance for pediatric gait rehabilitation,”
IEEE International Conference on Rehabilitation Robotics, pp. 332–339,
form to help children with Cerebral Palsy discover how to 8 2017.
walk”, reference PID2019-105110RB-C31, and the STRIDE [18] X. Chen, C. Ragonesi, J. C. Galloway, and S. K. Agrawal, “Training
project “VALIDACION TECNICA Y ECONOMICA DE toddlers seated on mobile robots to drive indoors amidst obstacles,”
IEEE Transactions on Neural Systems and Rehabilitation Engineering,
LA PLATAFORMA ROBOTICA DISCOVER2WALK PARA vol. 19, no. 3, pp. 271–279, 2011.
EL ENTORNO CLINICO”, reference PDC2022-133898-C31, [19] N. Smania et al., “Improved gait after repetitive locomotor training in
funded by the Spanish Ministry of Science and Innovation. children with cerebral palsy,” American Journal of Physical Medicine
& Rehabilitation, vol. 90, no. 2, pp. 137–149, February 2011.
This work was founded by the training program fellowship, [20] G. Delgado-Oleas, P. Romero-Sorozabal, J. Lora-Millan, A. Gutier-
PRE2020-092049, from the Spanish Ministry of Science and rez, and E. Rocon, “Bioinspired hierarchical electronic architecture
Innovation and the Juan de la Cierva program, with reference for robotic locomotion assistance: Application in exoskeletons,” IEEE
Access, pp. 1–1, 2023.
FJC2021-046692-I, granted by of the Spanish Ministry of [21] N. Hogan, “Impedance Control: An Approach to Manipulation: Part
Science and Innovation. I—Theory,” Journal of Dynamic Systems, Measurement, and Control,
vol. 107, no. 1, pp. 1–7, 03 1985.
R EFERENCES [22] P. Romero-Sorozábal, G. Delgado-Oleas, A. F. Laudanski, Á. Gutiérrez,
and E. Rocon, “Novel methods for personalized gait assistance: Three-
[1] S. Jezernik, G. Colombo, T. Keller, H. Frueh, and M. Morari, “Robotic dimensional trajectory prediction based on regression and lstm models,”
orthosis lokomat: A rehabilitation and research tool,” Neuromodulation: Biomimetics, vol. 9, no. 6, 2024.
Technology at the Neural Interface, vol. 6, no. 2, pp. 108–115, 2003. [23] A. Pott, “An algorithm for real-time forward kinematics of cable-driven
[2] J. F. Veneman et al., “Design and evaluation of the lopes exoskeleton parallel robots,” Advances in Robot Kinematics: Motion in Man and
robot for interactive gait rehabilitation,” IEEE Transactions on Neural Machine, pp. 529–538, 2010.
Systems and Rehabilitation Engineering, vol. 15, no. 3, pp. 379–386, [24] A.Pott, “Cable-driven parallel robots: Theory and application,” Springer
2007. Tracts in Advanced Robotics, vol. 120, pp. 1–465, 2018.
[3] S. Hesse and D. Uhlenbrock, “A mechanized gait trainer for restoration [25] J. Muñoz, P. R. Sorozabal, and E. R. de Lima, “Modelo dinámico de un
of gait,” Journal of rehabilitation research and development, vol. 37, sistema de actuación por cables para control de pelvis en rehabilitación
pp. 701–708, 2000. de la marcha,” Jornadas de Automática, p. 45, 7 2024.
IEEE/ASME TRANSACTIONS ON MECHATRONICS, 9

VII. B IOGRAPHY S ECTION Alvaro Gutierrez received the M.Sc. degree in

electronic engineering and a Ph.D. degree in com-
puter science from the Universidad Politécnica de
Madrid (UPM), Madrid, Spain, in 2004 and 2009,
respectively. He was a Visiting Researcher with
the Institut de Recherches Interdisciplinaires et de
Développements en Intelligence Artificielle, Univer-
sité Libre de Bruxelles, Brussels, Belgium, in 2008.
He is currently an Associate Professor of Automatic
Pablo Romero-Sorozabal completed his undergrad- and Control Systems with the E.T.S. Ingenieros
uate degree in Industrial Electronic Engineering and de Telecomunicación, UPM. He has authored or
Automation at Carlos III University of Madrid in coauthored four book chapters and more than 100 refereed journal articles and
2019, followed by a master’s degree in Biomedi- conference proceeding papers. His research interests include swarm robotics,
cal Engineering at the Universidad Politécnica de control systems, sensor networks, haptic systems, robotics and demand-side
Madrid in 2020. During his academic career, he management applications.
contributed to the development of mobile robotic
devices at the startup company StarRobotics. While
pursuing his master’s degree, he collaborated with
the Neural and Cognitive Engineering Group at the
Center for Automation and Robotics (CSIC-UPM) Eduardo Rocon was born in Vitoria, Brazil, in
on projects related to low limb robotic rehabilitation. In 2021, he joined 1979. He received a degree in electrical engineering
the aforementioned group to commence his Ph.D. studies at the Universidad from the Universidade Federal do Espirı́to Santo
Politécnica de Madrid. His research focuses on the development of human- (UFES) in 2001 and a Ph.D. degree from the Uni-
robot interaction strategies for the robotic rehabilitation of children with versidad Politécnica de Madrid in 2006. He is cur-
cerebral palsy. rently a researcher at the Spanish National Research
Council. He leads a research group.

Gabriel Delgado-Oleas was born in Cuenca,

Ecuador, in 1990. He obtained his B.S. in Electronic
Engineering from the University of Azuay in 2014
and his M.S. in Automation and Robotics from the
Polytechnic University of Madrid, Spain, in 2015.
He is currently pursuing his Ph.D. in Automatic
and Robotics at the same institution. Gabriel has
served in both teaching and research roles at the
University of Azuay. Additionally, he collaborates
with the Neural and Cognitive Engineering Group at
the Consejo Superior de Investigaciones Cientı́ficas.
His research interests include Robotics and Rehabilitation, Bioengineering,
Control Architectures, and Electronic Systems. His research interests include
developing robotic and electronic systems to assist neurologically impaired
subjects, especially their control algorithms and the interaction between users
and devices.

Jorge Muñoz Yañez-Barnuevo Jorge Muñoz re-

ceived the PhD in Electrical Engineering, Electronics
and Automation in 2020 from the University Carlos
III of Madrid after an MSc in Robotics and Automa-
tion. During this period, the research was focused on
several projects, among others, Handle (7th frame-
work), the Humanoid Robot TEO (RH2), and several
soft robotics projects (HumaSOFT, SofIA), all from
a robust and adaptive control point of view, including
techniques such as fractional control and system
identification. At present, within the framework of a
Juan de la Cierva postdoctoral program, the current line of research is evolving
towards the application of advanced control techniques to new platforms,
such as low-cost robotics and exoskeletons. He is currently working on the
control of the discover2walk device (cable-actuated exoskeleton for walking
rehabilitation), and starting a line of research in soft robotics systems applied
to bipedal walking.