Kinematic Validation of a Robotic Exoskeleton for Assisting Seated

Pelvic Movements by Wheelchair Users with Trunk Impairments

Chawin Ophaswongse1 , Victoria Lent2 , and Sunil K. Agrawal1

Abstract— Individuals with severe impairments in trunk shifts and relieve pressure between the sitting surface and
control have difficulty maintaining and controlling upright the skin, which is essential to prevent pressure injuries
posture during static or dynamic sitting. Passive devices like [4]. Upper extremity pain is also a common issue among
straps and orthoses may be used currently to support the trunk
in a wheelchair at the expense of user mobility. Alternatively, individuals with SCI due to overuse of the upper limb to
we propose that an active support to the pelvic segment may compensate for the lack of trunk control during functional
enhance daily activities while maintaining stability on the seat. activities, including pressure relief maneuvers, transfers, and
This paper presents a kinematic validation of a fabricated wheelchair mobility [5], [6].
robotic device for pelvic assistance on a wheelchair, and we While sitting, motion of the pelvis plays an important
name this device as a pelvic Wheelchair Robot for Active
Postural Support (pWRAPS). The robot has an in-parallel role during the functional movements in daily activities. The
architecture with three rotational degrees-of-freedom (DOFs). control of the trunk often starts at the neck and progresses
The end-effector translation is also coupled with the rotation to down to the pelvis. For someone with poor trunk control,
accommodate the natural movement of the pelvis on the seat. the lower segments of the trunk need to be stabilized in
The device was tested with seven healthy subjects to evaluate order to effectively use the upper segments. Therefore, our
the workspace provided by the device compared to their natural
range of motion (ROM). The position accuracy of the device proposed solution is to build a pelvic robot module that can
was also validated against a virtual reality (VR) motion tracking dynamically support natural range of motion (ROM) of the
system. pelvis. The system needs to accommodate a range of pelvis
heights, widths, and tilt angles in different directions on the
I. I NTRODUCTION
seat.
The ability to remain stable in an upright sitting position Our previous work has presented a novel methodology to
and return back to this body position after performing design and optimize the pelvic Wheelchair Robot for Active
a dynamic movement is an essential skill for performing Postural Support (pWRAPS) [7]. The robot architecture is a
activities of daily living (ADLs) [1]. However, people with fully-actuated 3-DOF parallel mechanism that accommodates
specific neurological disorders lack synergistic control of key an individual’s pelvic seated ROM and instantaneous screw
postural muscles as well as integrated sensory inputs in these axes (ISAs) of the pelvic segment from motion capture data.
movements. This results in poor voluntary trunk control and In this paper, we present a first-of-its-kind fabricated
movement compensation by non-postural muscles [2]. prototype of the pWRAPS and evaluate the device workspace
According to the National Spinal Cord Injury Statistical with able-bodied subjects and its kinematic accuracy against
Center, the number of people in the U.S. in 2021 living with a reference motion tracking system. The manuscript is orga-
spinal cord injury (SCI) is estimated to be approximately nized as follows: First, the fabricated hardware, mathematical
296,000 persons, with a range from 252,000 to 373,000 per- models, and its control system are explained. Next, a protocol
sons. Of those, 59.8% have complete or incomplete injuries of the human subject experiment is described. Finally, we
in the cervical spine, which leads to tetraplegia and impaired present results and discussions of the data regarding the
trunk control [3]. Many of these individuals use a wheelchair performance of the pWRAPS prototype.
as their primary means of mobility and often face difficulties
in performing daily activities, such as eating, working on a II. ROBOTIC D EVICE
desk, opening a door, and reaching for an object, due to A. Hardware
limited trunk control and stability in the seated position. A CAD of the pWRAPS prototype with a user and its
Impaired trunk muscle control and the lack of sensory kinematic model is shown in Fig. 1 and the real hardware
feedback also greatly limit the ability to perform weight configuration with a human subject is shown in Fig. 2. The
1 Chawin robot [RRR]U-2[RR]S structure is mounted on a custom-
Ophaswongse and Sunil K. Agrawal are with
the Robotics And Rehabilitation Laboratory (ROAR Lab), made bench. The axes of the actuated revolute joints (R)
Department of Mechanical Engineering, Fu Foundation School are fixed relative to the bench coordinate frame {Ob }. The
of Engineering and Applied Science, Columbia University, moving coordinate frame {O p } is rigidly attached to the
New York, NY 10027, USA. [email protected];
[email protected] pelvic brace. The positions of the universal joint center (U)
2 Victoria Lent is with Rehabilitation and Regenerative Medicine of chain 1 and the spherical joint centers (S) of chains 2 and
at Columbia University College of Physicians & Surgeons, New 3 on the pelvic brace are fixed relative to {O p }. The actuated
York, NY 10032, USA, and Spinal Cord Injury Rehabilitation
at Helen Hayes Hospital, West Haverstraw, NY 10993, USA. joints (R) are driven by three maxon permanent magnet
[email protected] brushless DC motors with Hall sensors (EC60 647692,
  T
ith limb are denoted by p bi = bx,i by,i bz,i . Each of the
Brushless three constraint planar 3D surfaces imposed by the kinematic
DC Motor
Pelvic Brace limbs is defined by two T variables for a unit normal vector
bn = n n n and a scalar value ρi for determining
i x,i y,i z,i
a point on the surface. A position
 on the plane
T with respect
to the base frame is b di = dx,i dy,i dz,i = ρi b ni + b di,o ,
HTC Vive Tracker
where b di,o is an optional constant offset position from
the origin of the base frame. Since the spherical/universal
joint center always lies in the planar surface, the constraint
equation is
Chain 2: [RR]S  
࢔ଶ b
ܴ p p + b R p p bi − b di · b ni = 0. (1)
ܴ ࢔ଷ
ܴ Rewrite (1) as
Chain 1: [RRR]U
࢔ଵ
ܴ ܵ
ܴ nx,i (x p + ex,i ) + ny,i (y p + ey,i ) + nz,i (z p + ez,i ) − ρi = 0, (2)
ܷ ‫ݖ‬௣
 T
{ܱ௣ }
‫ݕ‬௣ where ex,i ey,i ez,i = b R p p bi − b di,o .
‫ݖ‬௕
‫ݔ‬௣ b p can be solved from the system of three linear equations
ܴ ܵ p
‫ݔ‬௕ ‫ݕ‬௕ ܴ
{ܱ௕ } in (2), given that all three unit normal vectors of the plane
Chain 3: [RR]S
are linearly independent.
2) Limb Inverse Kinematics: Once we have computed
the coordinate transformation of the platform from section
II-B.1, the rotation axis of the last joint in the spheri-
Fig. 1. CAD Model of pWRAPS prototype with kinematic model of
[RRR]U-2[RR]S architecture. cal/universal joint can be obtained with respect to the first
revolute joint coordinate in the ith limb {Oi }, of which the
z-axis is coaxial to the first joint. Then the joint angles in
maxon precision motors, Taunton, MA, USA) which are each limb case can be computed as follows:
connected to the rotary shafts mounted on the fixed bearings 1. An [RR]S chain is fully defined by a constant three
on the bench frame. Each motor has a built-in encoder (MILE Denavit-Hartenberg (DH) parameters (a1 , d2 , a2). Given a
T
421988) with 4,096 counts per turn (CPT), and a Ø52 mm spherical joint center position i ps,i = xs ys zs , the five
planetary gearhead (GP 52 C 223095) with a gear ratio joint angles (θ1 , θ2 , θ3 , θ4 , θ5 ) in DH parameters can be
of 113:1 with 30 N-m maximum output torque. The robot solved below
linkages consist of CNC-machined aluminum parts, ABS xs2 + y2s + z2s − a21 − a22 − d22
3D printed parts (Ultimaker S5, Ultimaker, Framingham, cos θ2 = := D
MA, USA), and off-the-shelf rotary bearings and ball joints.   2a1 a2 
In order to accommodate different subject sizes, linkage θ2 = arctan 2 ± 1 − D2 , D
dimensions can also be adjusted by arrays of clearance holes θ1 = arctan 2 (ys , xs ) − arctan 2 (a2 sin θ2 , a1 + a2 cos θ2 ) (3)
on some machined aluminum plates. In addition, four rotary
θ3 = arctan 2 (r23 , r13 ) − θ1 − θ2
quadrature incremental encoders with 3,600 pulses per revo-  
lution (CALT GHS38-06G3600BML5, Shanghai QIYI Elec- π
θ4 = arctan 2 r33 , 1 − r33 2 +
trical & Mechanical Equipment, Baoshan District, Shanghai, 2
China) are connected to the passive revolute joints (R). This θ5 = arctan 2 (−r32 , r31 )
simplifies the forward kinematic computation, where the
where rm,n are the entries of the rotational matrix of the last
center positions of the universal joint (U) and the spherical
joint coordinate frame with respect to the first rotary joint
joints (S) are calculated directly from the seven encoder
frame of the limb. Note that θ1 and θ2 are angles of the first
readings. The position and orientation of {O p } with respect
actuated revolute joint (R) and the passive revolute joint (R)
to {Ob } can be estimated by the least-squares rigid motion
in the planar subchain respectively, and θ3 , θ4 , and θ5 are
using SVD in [8] by using the three positions of the last
the three angles of the spherical joint (S). These angles are
joints of the three kinematic limbs.
required for constructing velocity Jacobian of the robot.
2. An [RRR]U chain is defined by four constant
B. Mathematical Models
DH parameters (a1 , a2 , d3 , a3 ), and all the joint positions
1) Platform Inverse Kinematics: Consider the moving (θ1 , θ2 , θ3 , θ4 , θ5 ) need to be solved in order to fully describe
pelvic brace coordinate frame {O p } having an origin position the twist system of this limb. Let rm,n denote an element in
bp = x
 T
p p yp zp and its orientation matrix b R p with the rotational matrix of the last rotational axis in the universal
respect to the inertial base frame {Ob }. Three known fixed joint with respect the first revolute joint, defined by the DH
local positions of spherical or universal joint centers of the convention, along with the position of the universal joint
 Disturbance
ܑ (Wearer)
Motor Torque Motor Current
Desired ‫܅‬ௗ௘௦ ՜ ૌௗ௘௦ Constant Feedback Output
Force-Moment ૌௗ௘௦ ܑௗ௘௦
Emergency Solve ૚ െ pWRAPS Force-Moment
Rotary ‫܅‬ Joint Torque ࡷ࣎ (Pelvic Brace) ‫܅‬
Encoder Stop Button ௗ௘௦
Gravity Compensating
۸ ܏ሺ‫ܙ‬ሻ
Timing Belt Torques
Brushless Jacobian Forward ‫ܙ‬
DC Motor Transmission Kin. Joint Position
Feedback
VR VR PC
Tracker Fig. 4. Force controller of pWRAPS with gravity compensation
Control PC
Custom-made
Bench Pelvic Brace
C. Control System
The overview of the control system is shown in Fig. 3.
Hardware interfaces between the sensors and the control
Fig. 2. pWRAPS hardware with a subject
PC (Ubuntu 16.04.7 LTS, Xenial Xerus) are carried out by
VR PC
using two Sensoray 826 PCI express analog and digital I/O
Tracker Poses pWRAPS Prototype data acquisition (DAQ) cards (Sensoray Co., Inc., Tigard,
ƒ Calibrate “Home” Configuration
OR, USA). One of the DAQ cards is connected to three
Validate Forward Kinematics
ƒ
ƒ Track Body Segment Kinematics HTC Vive Trackers
4 ൈ Rotary
ESCON 70/10, 4-Q Servo Controllers (maxon precision
ROSBridge
TCP Control PC
Quadrature
Incremental Encoder
motors, Taunton, MA, USA). The servo controllers are
3 ൈ maxon Brushless DC
Motor with Gearbox and
Encoder
powered by a 48 VDC power supply. Each of the servo
High-Level Ctrl
DC Current/
Feedback
controllers is configured by the ESCON Studio Software
ƒ Fwd./Inv. Kinematics
ƒ Dynamic Modeling Hardware Interfaces
(ver 2.2) to receive enable/disable digital signal from a DIO
ƒ Force/Trajectory
Planning pin of the DAQ card. The servo controller also receives an
ros_control

Low-Level Ctrl
3 ൈ ESCON 70/10
Servo Controller
analog set value (-10 to 10 VDC) from the DAQ card to
Sensoray 826
ƒ Joint Position/Velocity
ƒ Joint Effort
PCIe DAQ Card
regulate the current/torque of the DC motor (-2.5 to 2.5 A :
-31.82 to 31.82 N-m at the gearbox shaft). In addition, the
Fig. 3. Control system overview of pWRAPS DAQ board receives two differential analog signals from the
servo controller (-4.0 to 4.0 VDC), which indicate the actual
current/torque and the actual speed of the motor output shaft
 T
(-5.58 to 5.58 rad/s at the gearbox shaft). In addition, seven
center i pu,i = xu yu zu the joint angles are computed
in the following steps: counter channels on the DAQ cards receive pulse signals
from the incremental encoders of the three maxon motors
Assuming, sin θ4 ≥ 0, and the four passive revolute joints. The pulse values are
 converted to angular displacements of the joint shafts in
θ4 = arctan 2 1 − r33
2 , −r
33 , θ5 = arctan 2 (−r32 , r31 ) radians.
In this study, a force controller of the pWRAPS was im-
θ1 + θ2 + θ3 := θ123 = arctan 2 (r23 , r13 ) plemented in the experiment. The controller sends the feed-
x p = xu − a3 cos θ123 , y p = yu − a3 sin θ123 forward torque/current commands to the ESCON servo con-
x2p + y2p + z2u − a21 − a22 − d32 trollers with the built-in closed-loop controllers of the motor
cos θ2 = := D currents, as shown in Fig. 4. A forward-backward Newton-
  2a1 a2  Euler inverse dynamics algorithm [9] is implemented to
θ2 = arctan 2 ± 1 − D2 , D estimate motor torques required to compensate the weights
θ1 = arctan 2 (y p , x p ) − arctan 2 (a2 sin θ2 , a1 + a2 cos θ2 ) of the linkages in each limb throughout the workspace of the
pWRAPS.
θ3 = θ123 − θ1 − θ2
 The control PC also communicates with the virtual reality
θ4 = arctan 2 1 − r33
2 ,r
33
(VR) PC which is connected to a HTC Vive VR headset,
base stations, controllers, and trackers (HTC, Taoyuan, Tai-
θ5 = arctan 2 (−r32 , r31 ) . wan). The Unity Real-Time Development Platform (version
(4) 2020.3.2f1, Unity Technologies, San Francisco, CA, US) is
used to interface with the VR devices. The tracker positions
Note that there are elbow down-up solutions (+θ2 , −θ2 ) in and orientations (poses) are streamed to the control PC
both kinds of the kinematic limbs. The solutions of the elbow in real-time by using the rosbridge TCP communication
positions are selected based on minimizing hardware inter- protocol in the ROS# (ros-sharp) package (Siemens AG,
ference with the wearer. In addition, the inverse kinematic Munich, Germany).
solution of every limb must exist, otherwise the desired end- The streamed tracker orientations on the bench and on
effector pose is considered unfeasible. the pelvic brace are essential for calibrating the home
 TABLE I
S EATED MOVEMENTS IN P WRAPS EXPERIMENT

No. Movement Description

Move the torso to perceived extreme ROM in a
1-6 plane of motion (sagittal plane: flexion-extension,
frontal plane: right/left lateral bending, and transverse
plane: right/left axial rotation) while still maintaining
“Free”: Neutral pWRAPS: Neutral
stability. Pause for 2-3 seconds then return to the
(a) (b)
neutral position.
Six cycles between the extreme ROM in each of the
7-9 three planes of motion. The motions were paced by a
metronome at 120 beats/sec, four beats long between
the neutral to an extreme position.
Six cycles of rolling the torso around the vertical axis
10 in the counter clockwise (CCW) direction, paced by
a metronome at 120 beats/sec. Each quadrant of a
cycle was four beats long.
“Free”: Flexion ሺെ߰௫ ሻ pWRAPS: Flexion ሺെ߰௫ ሻ
(c) (d)

pared to pROM when the pelvic brace is not attached

to the mechanism (“free” condition), and
2) to validate the forward kinematics computation against
the actual pelvic brace position in the space measured
by virtual reality (VR) trackers.
“Free”: R Lateral Bending ሺ൅ߠ௬ ሻ pWRAPS: R Lateral Bending ሺ൅ߠ௬ ሻ The experiment was approved by the Institutional Re-
(e) (f)
view Board (IRB) of Columbia University (Protocol No.
AAAR7388). Seven able-bodied adult subjects (4 males and
3 females) participated in the study with an average age of
24 years old (SD = 4.8 yr., range: 18 - 32 y/o.), 68.1 kg
average weight (SD = 13.4 kg, range: 45.5 - 86.4 kg), and
172.7 cm average height (SD = 10.1 cm, range: 155 - 184
cm). The average torso height is 50.6 cm (SD = 7.2 cm,
“Free”: L Axial Rotation ሺ൅߶௭ ሻ
range: 39.4 - 61.0 cm) measured from the greater trochanter
pWRAPS: L Axial Rotation ሺ൅߶௭ ሻ
(g) (h)
level to the suprasternal notch. The recruited subjects were
within the adjustable range of the pelvic brace sizes.
Fig. 5. A subject performs different seated movements while wearing the The subjects performed the same set of torso movements
pelvic brace in “free” and pWRAPS conditions
while wearing the pelvic brace with and without the mecha-
nism attached to the body. The movements are described in
configuration of the pWRAPS, which is the transformation Table I. The first six movements are static single plane ROM.
matrix of the pelvic brace coordinate frame {O p } from the The next three movements are the cyclical motions in each of
bench coordinate frame {Ob }. The home configuration is the planes. Finally, movement No. 10 is a continuous rolling
where all the encoder readings are at zero positions. To set around the vertical axis in the counter-clockwise (CCW)
a home configuration, when the pelvic brace is moved to a direction, which is a combination of motions in different
desired position, a rotational matrix b R p is calculated from planes. In the pWRAPS condition, the robot is controlled in
the current poses of the two trackers. Next, the obtained the transparent mode (Wdes = 0) with feed-forward gravity
rotational matrix is used to calculate all the absolute joint compensation torques to the motors.
angles by the inverse kinematics in section II-B. While the The “free” and the pWRAPS conditions at neutral config-
data acquisition is running, the incremented readings from urations are shown in Fig. 5 (a) and (b) respectively. Some
the encoders are added to the home absolute joint angles, and selected motions in different planes were performed by a
then these current absolute joint angles are used to calculate subject in both of the conditions, shown in Fig. 5: flexion
the forward kinematics in real-time. (sagittal plane) in (c) and (d); right lateral bending (frontal
plane) in (e) and (f), and left axial rotation (transverse plane)
III. H UMAN E XPERIMENT
in (g) and (h). The position and orientation of the pelvic
The pWRAPS prototype was evaluated with able-bodied brace were recorded by an attached VR tracker (100 Hz).
subjects to Additionally, forward kinematic solutions of the pelvic brace
1) characterize the pelvic range of motion (pROM) when at each time step (50 Hz) were recorded during the pWRAPS
wearing the pWRAPS in the transparent mode com- condition.
 TABLE II
G ROUP AVERAGE %ROM OF P WRAPS COMPARED TO “ FREE ”
CONDITION

%pROM Flexion Extension R Lateral Bend

(−ψx ) (+ψx ) (+θy )
MEAN 36.19% 33.05% 85.35%
SD 17.48% 12.01% 33.05%
%pROM L Lateral Bend R Axial Rotation L Axial Rotation
(−θy ) (−φz ) (+φz )
MEAN 87.44% 73.96% 63.51%
SD 10.78% 24.59% 18.72%

IV. R ESULTS AND D ISCUSSIONS

A. Pelvic Range of Motion
The peak rotational angles in the primary planes of motion
(Movement Nos. 1-6 in Table I) were used to evaluate the (a)
percent pelvic range of motion (%pROM) when wearing
the pWRAPS compared to the “free” pelvic movement
condition. The average %pROM of the pWRAPS is shown
in Table II, where data points that extend over 1.5 times
of interquartile range (IQR = Q3 - Q1) from the edges
of the box plots were excluded as outliers. Overall, the
group average of the %pROM in the lateral bending direction
(θy ) are over 85% in both directions, which are the largest
percentages among all the rotational directions. The average
%pROM in the axial rotation are over 60%. The flexion and
extension angles are shown to have smaller %pROM than
other single plane ROMs at 36.19% and 33.05% respectively.
This may be due to exaggerated flexion and extension angles
from sliding motions of the pelvic brace relative to the torso
of the subject during the “free” condition.
(b)
B. Kinematic Validation
The data from the cyclical movements in single planes of Fig. 6. Rotational angles of the pelvic brace measured by the Vive tracker
motion (Movement Nos. 7-9) and combination of different vs. forward kinematics from the encoder readings during cyclic movements
planes (Movement No. 10) in Table I were used to evaluate by a selected subject: (a) flexion-extension and (b) lateral bending.
the accuracy of the forward kinematic calculation from the
encoder readings described in Section II-A. A subject was
excluded from the data analysis due to a malfunction of which created some deflection in the mechanical components
the tracking system. The results of group average (n = and thus increased the tilt angles of the pelvic brace beyond
6) of mean absolute error (MAE), root-mean-square error the encoder readings.
(RMSE), and peak of absolute error (Peak) over each of According to the results from the experiment, the mech-
the movement sessions are shown in Table III. Also, the anism can accommodate over 60 %pROM on an average
recorded rotational angles of the pelvic brace over time across a group of able-bodied adults in the lateral bending
(VR tracker vs. forward kinematics) in the four movement and axial rotation, which seems promising for the pressure
sessions performed by a selected subject are shown in Fig. relief applications. The small %pROM observed in the flex-
6 and Fig. 7. ion and extension angles may be due to the sliding motion
The group average MAE and RMSE of the main angle in of the pelvic brace during the “free” condition. Friction in
each of the planes are under 4.0 degrees, while the maximum the mechanism may also inhibit some motion in the sagittal
average peak is at 8.5 degrees in the yaw (φz ) direction. All plane. This problem can be overcome by adding friction
the average MAE and RMSE values in the rolling motion compensation torque commands to the motor or by using
are also under 4.0 degrees, while the maximum average peak other control strategies in the transparent mode that are more
error is 8.1 degrees in the roll (θy ) direction. responsive to an external force-moment applied by the user.
In Fig. 6 (a) and 7 (b), the errors in the flexion-extension Ball joints may also limit the workspace of the device and
angle (pitch, ψx ) are increased around the peaks of the cycles. can be replaced by ones with larger range of motion.
This may be due to the external gravitational moment from The accuracy of the position sensors of the pWRAPS
the human body at the extreme positions in the sagittal plane, against the validating VR tracking system are under 4
 at extreme positions due to deflections of the mechanical
components are kept below 9 degrees. We believe this
level of accuracy is acceptable for postural training and
rehabilitation applications.
Also, the pelvic brace design and the mechanism archi-
tecture can adapt to a wide range of subject sizes. The
pelvic brace can fit a waist circumference ranging from
71.1-96.5 cm, which translates to pant sizes XS to L (US).
The adjustability of the linkage sizes in the [RRR]U-2[RR]S
architecture makes it possible for the subject to comfortably
wear the pelvic brace in their neutral configurations.
V. C ONCLUSIONS
This paper presents the first fabricated prototype of the
pWRAPS. The device wearability and workspace were eval-
uated with able-bodied subjects. On the average, over 85%
(a) of lateral bending, over 60% of axial rotation, and over 30%
of flexion-extension, compared to the natural ROM of the
pelvis, are provided by the device. The device kinematic error
becomes larger at the edge of the workspace.
Future works include development of a task-based postural
training using an active force controller and visual feedback
to evaluate the efficacy of the device for rehabilitation. An
intuitive user-based control of the device will be developed
for users who lack trunk mobility. The hardware can be
modified to improve physical interfaces with the users and
to accommodate average anthropomorphic dimensions of the
study population before clinical trials. Finally, next iterations
of the pWRAPS prototype will be focused on reducing size,
weight, and complexity, and improving the feasibility of
integrating the device to an actual wheelchair.
(b) ACKNOWLEDGMENT
This work is supported by the New York State Innovative,
Fig. 7. Rotational angles of the pelvic brace measured by the Vive tracker Developmental or Exploratory Activities (IDEA) in Spinal
vs. forward kinematics from the encoder readings during cyclic movements
by a selected subject: (a) axial rotation and (b) rolling in counter-clockwise
Cord Injury contract no. C35595GG.
direction.
R EFERENCES
TABLE III [1] B. Gagnon, C. Vincent, and L. Noreau, “Adaptation of a seated
G ROUP AVERAGE (±SD) OF ROTATION MEASUREMENT ACCURACY BY postural control measure for adult wheelchair users,” Disability and
Rehabilitation, vol. 27, pp. 951–959, 9 2005.
P WRAPS (n = 6). B OLD VALUES ARE THE PRIMARY ANGLES IN THE [2] H. A. M. Seelen, Y. J. M. Potten, A. Huson, F. Spaans, and J. P. H.
PLANES OF MOTION . Reulen, “Impaired balance control in paraplegic subjects,” Journal of
Electromyography and Kinesiology, vol. 7, no. 2, pp. 149–160, 1997.
[3] National Spinal Cord Injury Statistical Center, “Spinal cord injury
Motion Error Pitch ψx (◦ ) Roll θy (◦ ) Yaw φz (◦ ) facts and figures at a glance,” University of Alabama at Birmingham,
Flexion- MAE 2.8 ± 0.7 1.9 ± 2.2 2.0 ± 0.6 Birmingham, AL, Tech. Rep., 2021.
Extension RMSE 3.4 ± 1.0 2.1 ± 2.2 2.1 ± 0.6 [4] N. Sezer, S. Akkuş, and F. G. Uğurlu, “Chronic complications of
Peak 7.6 ± 2.1 3.7 ± 3.4 3.5 ± 0.8 spinal cord injury,” World Journal of Orthopedics, vol. 6, no. 1, p. 24,
Lateral MAE 1.8 ± 1.2 2.8 ± 1.4 2.7 ± 1.0 2015.
Bending RMSE 2.0 ± 1.2 3.3 ± 1.6 2.9 ± 1.0 [5] J. V. Subbarao, J. Klopfstein, and R. Turpin, “Prevalence and impact
Peak 4.4 ± 2.3 7.7 ± 3.6 5.4 ± 1.4 of wrist and shoulder pain in patients with spinal cord injury,” The
Axial MAE 2.1 ± 1.3 2.4 ± 1.9 3.4 ± 0.9 Journal of Spinal Cord Medicine, vol. 18, no. 1, pp. 9–13, 1 1995.
Rotation RMSE 2.5 ± 1.5 2.8 ± 1.9 3.9 ± 1.0 [6] M. Dalyan, D. D. Cardenas, and B. Gerard, “Upper extremity pain after
Peak 5.4 ± 3.5 6.8 ± 2.9 8.5 ± 2.5 spinal cord injury,” Spinal Cord, vol. 37, no. 3, pp. 191–195, 1999.
Rolling MAE 2.7 ± 0.9 3.0 ± 2.1 2.6 ± 1.3 [7] C. Ophaswongse and S. K. Agrawal, “Optimal design of a novel
(Combined) RMSE 3.3 ± 1.1 3.6 ± 2.3 2.8 ± 1.2 3-DOF orientational parallel mechanism for pelvic assistance on a
Peak 7.0 ± 2.5 8.1 ± 4.0 5.2 ± 2.0 wheelchair: An approach based on kinematic geometry and screw
theory,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp.
3315–3322, 2020.
[8] O. Sorkine-Hornung and M. Rabinovich, “Least-squares rigid motion
using svd,” Computing, vol. 1, no. 1, 2017.
degrees of MAE and RSME group averages across all [9] K. M. Lynch and F. C. Park, Modern Robotics: Mechanics, Planning,
movements. The average peak errors, which tend to occur and Control. Cambridge University Press, 2017.