← Back To Work

MIRACLE

Lower-Limb Exoskeleton   

Mechanical Design, Dynamic Modeling & Sensorless Impedance Control 

Overview

MIRACLE is a government-funded interdisciplinary research project focused on developing robotic systems to assist individuals with walking impairments.

MECHANICAL DEVELOPMENT

As a member of the Robotics Team at Yonsei University's MLCS Lab, I contributed to the mechanical development of a pediatric lower-limb exoskeleton, from CAD design to fabrication and assembly.

DYNAMICS & CONTROL

Building on the physical prototype, I modeled the exoskeleton's dynamics and investigated sensorless impedance control. I subsequently designed a disturbance-observer-enhanced control architecture and evaluated its tracking performance in MATLAB simulations.

Explore the Hardware

PHYSICAL PROTOTYPE

Unloaded Walking Test 

On-Body Gait Demonstration 

01

From concept
to mechanism.

Exoskeleton CAD Assembly 

Mechanical hardware development; walker frame excluded.

Full-System CAD Assembly 
Joint Assembly View 1
View 2
View 3

From CAD
to hardware.

Early hardware fabrication and integration process for the MIRACLE prototype.

Integrated
Prototype 

The final assembled hardware, integrating the mechanical components into a functional prototype.

Assembled MIRACLE Prototype — Multiple Views 

02

From hardware
to equations.

CAD-Based Dynamic Modeling

I first modeled each exoskeleton leg as a two-link system representing the thigh and shank, with hip and knee joints.

Two-Link Dynamic Model 

Equation of Motion

M(q)q¨+C(q,q˙)q˙+G(q)+uext=τM(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) + u_{\mathrm{ext}} = \tau
Full Dynamic Model Derivation

Linkage Assemblies & Mass Properties

Using Creo, I extracted the mass, center of mass, and moment of inertia of each linkage from the mechanical design.

Linkage 1 Assembly 

Linkage 1 Mass Properties

Scroll to inspect

Linkage 2 Assembly

Linkage 2 Mass Properties

Scroll to inspect

Dynamic Modeling & Equations of Motion 

A. Inertia Terms

I then constructed a double-pendulum model and derived its dynamics using the Euler–Lagrange formulation.
Based on the Jacobian matrices of each link's center of mass and their rotational inertias, I obtained the following inertia matrix.

Equation 01 — Jacobian-Based Inertia Matrix
M(q)=m1Jv1TJv1+m2Jv2TJv2+[I1+I2I2I2I2]M(q) = m_1 J_{v1}^{T} J_{v1} + m_2 J_{v2}^{T} J_{v2} + \begin{bmatrix} I_1 + I_2 & I_2 \\ I_2 & I_2 \end{bmatrix}
Equation 02 — Inertia Matrix
M(q)=[M11M12M21M22]M(q) = \begin{bmatrix} M_{11} & M_{12} \\ M_{21} & M_{22} \end{bmatrix}
Equation 03 — Expanded Matrix Elements
M11=m1lc12+m2(l12+lc22+2l1lc2cos⁡q2)+I1+I2M_{11} = m_1 l_{c1}^{2} + m_2 \bigl(l_1^{2} + l_{c2}^{2} + 2 l_1 l_{c2} \cos q_2\bigr) + I_1 + I_2
M12=M21=m2(l1lc2cos⁡q2+lc22)+I2M_{12} = M_{21} = m_2 \bigl(l_1 l_{c2} \cos q_2 + l_{c2}^{2}\bigr) + I_2
M22=m2lc22+I2M_{22} = m_2 l_{c2}^{2} + I_2
B. Coriolis & Centrifugal Terms

Using the inertia matrix, I derived the Coriolis and centrifugal matrix through the Christoffel symbols of the first kind.
This matrix describes the velocity-dependent forces and dynamic coupling between the hip and knee joints.

Equation 04 — Christoffel Symbols
Γijk=12(∂Mij∂qk+∂Mik∂qj−∂Mjk∂qi)\Gamma_{ijk} = \frac{1}{2} \left( \frac{\partial M_{ij}}{\partial q_k} + \frac{\partial M_{ik}}{\partial q_j} - \frac{\partial M_{jk}}{\partial q_i} \right)
Equation 05 — Coriolis and Centrifugal Matrix
C(q,q˙)=[−hq˙2−h(q˙1+q˙2)hq˙10]C(q,\dot{q}) = \begin{bmatrix} -h\dot{q}_2 & -h(\dot{q}_1 + \dot{q}_2) \\ h\dot{q}_1 & 0 \end{bmatrix}
h=m2l1lc2sin⁡(q2)h = m_2 l_1 l_{c2} \sin(q_2)
C. Gravitational Terms

I then derived the gravitational vector from the potential energy of both links.
By differentiating the potential energy with respect to the joint angles, I obtained the gravitational torque acting on each joint.

Equation 06 — Gravitational Potential Energy
P(q)=−m1glc1cos⁡q1−m2g(l1cos⁡q1+lc2cos⁡(q1+q2))P(q) = -m_1 g l_{c1} \cos q_1 - m_2 g \bigl(l_1 \cos q_1 + l_{c2} \cos(q_1 + q_2)\bigr)
Equation 07 — Gravitational Vector
G(q)=∂P(q)∂qG(q) = \frac{\partial P(q)}{\partial q}
G1=m1glc1sin⁡q1+m2g(l1sin⁡q1+lc2sin⁡(q1+q2))G_1 = m_1 g l_{c1} \sin q_1 + m_2 g \bigl(l_1 \sin q_1 + l_{c2} \sin(q_1 + q_2)\bigr)
G2=m2glc2sin⁡(q1+q2)G_2 = m_2 g l_{c2} \sin(q_1 + q_2)
D. Complete Equation of Motion

Finally, I combined the inertia, Coriolis and centrifugal, and gravitational terms to obtain the complete equation of motion for the two-link exoskeleton.

This model describes the joint dynamics under applied and external disturbance torques and provided the foundation for developing the sensorless impedance controller and disturbance observer.

Equation 08 — Equation of Motion
M(q)q¨+C(q,q˙)q˙+G(q)+uext=τM(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) + u_{\mathrm{ext}} = \tau
03

From interaction
to robust control.

BASELINE

ROB-Based Sensorless Impedance Control 

For a wearable robot, following a predefined trajectory alone is insufficient. The robot must also respond to external forces generated through interaction with the wearer.

To address this, I implemented a sensorless impedance control framework based on a Reaction Torque Observer (ROB), following the approach proposed by Tran et al. Instead of relying on additional force/torque sensors, the ROB estimates external interaction torque using motor encoder feedback and the robot's dynamic model.

ROB-Based Sensorless Impedance Control Architecture (Adapted from Tran et al. https://doi.org/10.1109/AIM.2015.7222611) 

DESIGN MOTIVATION

Although this framework enables sensorless impedance control, its performance can be affected by uncertainties in the dynamic model.
During the development of the baseline controller, I observed a delay between the reference gait trajectory and the actual joint motion.
This motivated me to investigate an additional disturbance compensation method to improve tracking performance.

PROPOSED

Disturbance Observer-Enhanced Robust Control

To address this, I integrated a nonlinear Disturbance Observer (DO) into the existing ROB-based sensorless impedance control framework.

Unlike previous approaches that combine the two observers in parallel, I proposed a serial architecture in which the DO first estimates and compensates for disturbances and model uncertainties. The ROB then estimates the remaining external interaction torque using the compensated system dynamics.

My goal was to improve the accuracy of external torque estimation and reduce trajectory-tracking errors.

Proposed serial DO–ROB control architecture. The DO compensates for disturbances and model uncertainties before the ROB estimates the remaining external interaction torque.

Observer Design & Stability Analysis
A. ROB-Based Sensorless Impedance Control

Based on the derived dynamic model, I formulated the external torque estimation process under the model compensation assumptions.
However, directly differentiating encoder measurements can amplify noise.
To address this, the ROB uses a low-pass filter to estimate external interaction torque while reducing the effects of measurement noise.

Equation 09 — Reaction Torque Estimation
u^ext=gs+g(ua−M^(q)sq˙)\hat{u}_{\mathrm{ext}} = \frac{g}{s+g}\left(u_a - \hat{M}(q)s\dot{q}\right)

I then incorporated the estimated interaction torque into the desired impedance model.
By defining the desired inertia, damping, and stiffness, the controller can adjust the exoskeleton's mechanical response to external interaction.

Using the impedance tracking error, I derived the commanded acceleration and joint torque to achieve trajectory tracking while maintaining the desired impedance behavior.

Equation 10 — Desired Impedance Model
Zd(s)=Mds2+Bds+KdZ_d(s) = M_d s^2 + B_d s + K_d
Equation 11 — Impedance Tracking Error
eimp=qd−Zd−1u^ext−qe_{\mathrm{imp}} = q_d - Z_d^{-1}\hat{u}_{\mathrm{ext}} - q
Equation 12 — Impedance Error Derivative
e˙imp=q˙d−sZd−1u^ext−q˙\dot{e}_{\mathrm{imp}} = \dot{q}_d - s Z_d^{-1}\hat{u}_{\mathrm{ext}} - \dot{q}
Equation 13 — Commanded Acceleration
aq=q¨d+Kpeimp+Kve˙impa_q = \ddot{q}_d + K_p e_{\mathrm{imp}} + K_v\dot{e}_{\mathrm{imp}}
Equation 14 — Commanded Joint Torque
u=M^(q)aq+u^ext+C^(q,q˙)q˙+G^(q)u = \hat{M}(q)a_q + \hat{u}_{\mathrm{ext}} + \hat{C}(q,\dot{q})\dot{q} + \hat{G}(q)
B. Disturbance Observer Design

Disturbance Observer-Enhanced Robust Control

Starting from the previously derived equation of motion, I formulated the disturbance model and designed a nonlinear observer to estimate the disturbance torque.

To avoid directly using joint acceleration measurements, which can amplify encoder noise through numerical differentiation, I introduced an auxiliary vector.
This allowed me to formulate the observer using the robot's dynamic model and joint velocity information.

Equation 15 — Disturbance Model
d=M(q)q¨+C(q,q˙)q˙+G(q)−ud = M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) - u
Equation 16 — Auxiliary Vector
z=d^−p(q,q˙)z = \hat{d} - p(q,\dot{q})

I then designed the auxiliary function based on the joint velocities and derived the corresponding observer gain matrix using the inertia matrix.

Equation 17 — Auxiliary Function
p(q,q˙)=c[q˙1q˙1+q˙2]p(q,\dot{q}) = c\begin{bmatrix} \dot{q}_1 \\ \dot{q}_1 + \dot{q}_2 \end{bmatrix}
Equation 18 — Observer Gain Matrix
L(q)=c[1011]M(q)−1L(q) = c\begin{bmatrix} 1 & 0 \\ 1 & 1 \end{bmatrix}M(q)^{-1}

By substituting the auxiliary function into the observer formulation, I derived the observer dynamics without explicitly requiring joint acceleration measurements.

The estimated disturbance can then be reconstructed from the auxiliary vector and the designed function.
This estimate is used for disturbance compensation in the proposed controller.

Equation 19 — Disturbance Observer Dynamics
z˙=L(q)[C(q,q˙)q˙+G(q)−u−z−p(q,q˙)]\dot{z} = L(q)\bigl[C(q,\dot{q})\dot{q} + G(q) - u - z - p(q,\dot{q})\bigr]
Equation 20 — Estimated Disturbance
d^=z+p(q,q˙)\hat{d} = z + p(q,\dot{q})

To ensure that the disturbance observer could provide a reliable estimate, I further analyzed its estimation-error dynamics using Lyapunov theory.
This analysis allowed me to derive a sufficient condition for selecting the observer gain based on the mechanical parameters of the exoskeleton.

C. Stability Analysis

Lyapunov-Based Stability Analysis 

To examine the stability of the proposed disturbance observer, I first defined the disturbance estimation error as the difference between the actual and estimated disturbances.

Assuming that the disturbance remains approximately constant over the observer's estimation timescale, I derived the estimation-error dynamics.
This provided the foundation for analyzing whether the observer error converges to zero.

Equation 21 — Disturbance Estimation Error
e=d−d^e = d - \hat{d}
Equation 22 — Estimation Error Dynamics
e˙+L(q)e=0\dot{e} + L(q)e = 0

To analyze the estimation-error dynamics, I introduced a coordinate transformation of the inertia matrix.
This allowed me to express the observer dynamics in a form suitable for Lyapunov analysis.

Using the transformed inertia matrix, I constructed a positive-definite Lyapunov candidate function to examine the convergence of the disturbance estimation error.

Equation 23 — Inertia Matrix Transformation
Mˉ(q)=ATM~(q)A,A=[1011]\bar{M}(q) = A^T\widetilde{M}(q)A,\qquad A = \begin{bmatrix} 1 & 0 \\ 1 & 1 \end{bmatrix}
Equation 24 — Lyapunov Candidate Function
V(e,q)=eTMˉ(q)eV(e,q) = e^T\bar{M}(q)e

I then differentiated the Lyapunov function with respect to time and substituted the estimation-error dynamics.

By expressing its derivative as a quadratic form, I examined the conditions required for the derivative to remain negative definite.
This led to a sufficient observer gain condition based on the mechanical parameters and the maximum angular velocity of the second link.

Equation 25 — Lyapunov Derivative
V˙(e,q)=−eTQ(q,q˙)e\dot{V}(e,q) = -e^T Q(q,\dot{q})e
Equation 26 — Positive-Definiteness Condition
Q(q,q˙)=[2c−c+η−c+η2c]>0Q(q,\dot{q}) = \begin{bmatrix} 2c & -c + \eta \\ -c + \eta & 2c \end{bmatrix} > 0
η=m2l1lc2q˙2sin⁡(q2)\eta = m_2 l_1 l_{c2}\dot{q}_2\sin(q_2)
Equation 27 — Observer Gain Condition
c>m2l1lc2q˙2,max⁡c > m_2 l_1 l_{c2}\dot{q}_{2,\max}

The resulting condition relates the observer gain to the mass, geometry, and maximum angular velocity of the second link.

Under the assumptions used in the analysis, satisfying this condition provides a sufficient basis for asymptotic convergence of the disturbance estimation error.

This analysis allowed me to establish a theoretical criterion for selecting the observer gain rather than relying solely on empirical tuning.
The proposed controller was subsequently evaluated through MATLAB simulations.

04

Evaluated
in simulation.

MATLAB SIMULATIONS

Trajectory Tracking Performance

I evaluated the proposed control architecture in MATLAB by comparing the baseline ROB-only controller with the enhanced DO+ROB controller.

The animated plot compares the reference trajectory and the simulated joint response for the hip and knee joints. Compared with the ROB-only baseline, the DO+ROB controller followed the reference trajectory more closely, particularly at the knee joint.

This improvement is also reflected in the RMS tracking error. The enhanced controller reduced the RMS error by 14.7% at the hip joint and 27.4% at the knee joint, indicating improved trajectory-tracking performance in simulation.

RMS Tracking Error Comparison
JointROB OnlyDO+ROBReduction
Hip0.03410.029114.7%
Knee0.06860.049827.4%
3D Motion Simulation 
Animated comparison of trajectory tracking performance

Impedance Behavior under External Disturbances

To evaluate whether the proposed DO+ROB controller could maintain the desired impedance behavior under external disturbances, I conducted additional MATLAB simulations using two impedance settings: Soft and Stiff.

I applied simulated random disturbances and compared the joint responses under different damping and stiffness parameters while maintaining the same desired inertia. The Soft setting was designed to allow greater compliance, whereas the Stiff setting was intended to provide greater resistance to external disturbances.

Simulation parameters 

ParameterSoftStiff
Desired inertiakg·m²55
Desired dampingN·m·s/rad575
Desired stiffnessN·m/rad101000
Soft Impedance Response 
Stiff Impedance Response  

The results showed that the Soft setting allowed larger deviations from the reference trajectories, while the Stiff setting maintained closer tracking, particularly at the hip joint. This behavior was consistent with the expected relationship between mechanical impedance and disturbance response.

The RMS tracking errors further confirmed these differences. Overall, the simulations demonstrated that the proposed DO+ROB controller retained the expected impedance-dependent behavior under external disturbances.

RMS Tracking Error Comparison
JointSoftStiff
Hip0.08640.0240
Knee0.05370.0489
05

What I
contributed.

Mechanical Design & Hardware Development

I was responsible for the mechanical design of the lower-limb exoskeleton, excluding the walker frame. I developed the CAD models of its structural and joint components and collaborated with manufacturing partners to fabricate and assemble the physical prototype.

Dynamic Modeling & Sensorless Control

Using the CAD-derived physical parameters, I formulated the two-link dynamic model through the Euler–Lagrange method. Based on previous research, I also developed and evaluated a ROB-based sensorless impedance controller.

Robust Control & Stability Analysis

To address the observed trajectory-tracking delay, I proposed a serial DO–ROB control architecture and derived the corresponding disturbance observer equations. I analyzed the observer's estimation-error convergence using Lyapunov theory and evaluated the proposed controller through MATLAB simulations, achieving 14.7% and 27.4% reductions in RMS tracking error at the hip and knee joints, respectively.

06

A record
of development.

A visual record of MIRACLE's development, capturing the iterative process of mechanical design, component fabrication, assembly, and prototype integration.

07

The following materials provide additional context for the MIRACLE project, including its hardware development, dynamic modeling, and control research. The technical references also acknowledge the foundational methods used to develop and evaluate the sensorless impedance control framework.

Research Materials

1. Technical Research Report — Geonung Jo (2024)

Enhancing the Performance of Sensorless Impedance Control for Lower Limb Exoskeleton using Robust Control

Research Report

2. Related MIRACLE Publication — Park et al. (2025)

Validity and Test–Retest Reliability of an Independent Self-Balanced, Socially Integrated Gait-Assisted Robotic System for Hip and Knee Joint Kinematics in Rehabilitation.

Journal of Mechanics in Medicine and Biology, 25(10), 2540109.

Publication · Yonsei University Research Portal 

3. MIRACLE — Government-Funded Research Project

Research Center for the Development of MIRACLE, an Autonomous Walking Robot to Improve the Quality of Life and Social Integration of Children and Older Adults with Gait Impairments.

Funding Agency: Ministry of Science and ICT (MSIT), Republic of Korea
Lead Institution: Yonsei University
Research Program: Group Research Support Program

Project Information: https://app.rndcircle.io/project/d567189a-9dcf-53a1-8e0a-c3fecf2009b0

References

[1] Dhand, S., Singla, A., & Virk, G. S. (2016). “A Brief Review on Human-Powered Lower-Limb Exoskeletons.” Conference on Mechanical Engineering and Technology (COMET-2016), IIT (BHU), Varanasi, India.

[2] Chen, W. H., Ballance, D. J., Gawthrop, P. J., & O'Reilly, J. (2000). “A Nonlinear Disturbance Observer for Robotic Manipulators.” IEEE Transactions on Industrial Electronics, 47(4), 932–938.

https://doi.org/10.1109/41.857974

[3] Chen, W. H. (2004). “Disturbance Observer Based Control for Nonlinear Systems.” IEEE/ASME Transactions on Mechatronics, 9(4), 706–710.

https://doi.org/10.1109/TMECH.2004.839034

[4] Sariyildiz, E. (2023). “A Stability Analysis for the Reaction Torque Observer-Based Sensorless Force Control Systems.” 2023 IEEE International Conference on Mechatronics (ICM).

https://doi.org/10.1109/ICM54990.2023.10101940

[5] Spong, M. W., Hutchinson, S., & Vidyasagar, M. (2020). Robot Modeling and Control. 2nd ed., John Wiley & Sons.

[6] Tran, Q. V., Kim, S., Lee, K., Kang, S., & Ryu, J. (2015). “Force/Torque Sensorless Impedance Control for Indirect Driven Robot-Aided Gait Rehabilitation System.” 2015 IEEE International Conference on Advanced Intelligent Mechatronics (AIM), 652–657.

https://doi.org/10.1109/AIM.2015.7222611