23 March 2025
Summary
In my final year at Nanyang Technological University, I had the opportunity to work on an exciting and impactful project in the field of assistive robotics. The focus of my project was to enhance the performance of the Assistive Robotic Arm Extender (ARAE) by personalizing the human arm dynamics model. By estimating individual-specific anthropometric data, I aimed to improve the robot’s ability to provide accurate, stable, and user-specific support—ultimately reducing muscle activation and fatigue leading to increased comfort for users with upper limb mobility impairments.
Introduction
Mobility impairment of the upper limbs is a widespread issue affecting various populations, particularly those with neurological conditions such as stroke, cerebral palsy (CP), and multiple sclerosis (MS). These impairments can significantly impact daily activities and overall quality of life.
Previous work on ARAE introduced a model to estimate the user’s arm posture using only the robot’s internal sensors—removing the need for external wearable devices. However, this model relied on nominal human parameters, which don’t account for individual differences in arm mass or limb structure. These inaccuracies can affect the effectiveness of support, potentially leading to discomfort or fatigue.
To address this, my project developed a method to personalize the human model based on each user’s unique biomechanics—allowing ARAE to provide truly tailored assistance.
Methodology
The core idea was to estimate a user’s immeasurable anthropometric parameters (like arm mass and center of mass) by comparing actual torques measured during robot-assisted arm postures with theoretical torques calculated from a model. Here’s how it was done:
- Simulation with MuJoCo
- I built a 3D simulation of ARAE and a human model in the MuJoCo physics engine.
- By applying torque via a PID controller to reach different ADL-like positions, I validated that personalized parameters closely matched actual torque values.
- Human Testing - Personalization
- Ten healthy participants were recruited for testing.
- The torques required to maintain those positions were logged and used to estimate individual anthropometric values through optimization.
- Human Testing - sEMG Analysis
- I placed wireless sEMG sensors on six key upper limb muscles to measure muscle activation and fatigue during different activities (forward reach, lateral reach, drinking, and scooping).
- Activities were performed under three modes: no robot, ARAE with nominal model, and ARAE with personalized model.
- Features like mean absolute value (MAV) and median frequency (MDF) were extracted to analyze performance.
Results
The personalization of the human arm dynamics model led to several notable improvements in the performance of the Assistive Robotic Arm Extender (ARAE). One of the most significant outcomes was the improvement in torque estimation accuracy. When comparing the actual torque exerted by ARAE with the theoretical torque calculated using both nominal and personalized models, the personalized model consistently showed lower error—particularly for Motor 2 and Motor 3, which are primarily responsible for supporting and lifting the arm. This indicates that the personalized model aligns more closely with the user’s true biomechanical needs, enabling more precise support force computation.
Another important finding was the effect of personalization on system stability. In previous versions of ARAE, Motor 1 often exhibited torque fluctuations due to encoder noise and loose coupling with the user’s arm, which sometimes caused instability. With the personalized model, the torque output of Motor 1 was consistently close to zero across all positions. This helped suppress fluctuations, contributing to more stable and safer robot behavior, particularly during static support.
In terms of muscle activation and fatigue, results from surface electromyography (sEMG) revealed that personalization had a beneficial effect on extension muscles. Muscles such as the triceps brachii and posterior deltoid showed reduced activation and fatigue across all activities when using the personalized model compared to the nominal one. This means users experienced less physical effort and strain during tasks involving arm extension.
For flexion muscles, such as the biceps brachii and anterior deltoid, the outcomes were more mixed. While some users experienced slightly higher activation levels with the personalized model, the overall impact was relatively small. In most cases, the personalized model maintained or slightly reduced fatigue levels, even when muscle activation increased.
Overall, the results demonstrate that personalizing the human dynamics model significantly improves assistive performance, especially in tasks requiring arm extension. The improved torque accuracy and system stability make ARAE safer and more effective, while sEMG analysis confirms that tailored support helps reduce muscle workload in critical muscle groups.
Skills
ARAE project has allowed me to develop and refine a diverse set of skills, including:
- Biomechanical modeling – Understanding and applying human dynamics in robotic systems
- Robotic control systems – Implementing and tuning PID/PD controllers through ROS and C++
- Physics simulation (MuJoCo) – Developing and validating robotic models in simulation environments
- Data analysis – Processing and interpreting sensor and torque data using MATLAB
- Surface electromyography (sEMG) – Setting up, collecting, and analyzing muscle signal data
- Optimization techniques – Solving inverse problems using numerical methods
- Hardware integration – Calibrating, debugging, and testing a physical robot with human subjects
- Scientific writing – Producing technical documentation, reports, and publications
- Human-subject testing – Designing and conducting experiments ethically and effectively
Reflection
Working on this project has been one of the most challenging yet rewarding experiences of my undergraduate journey. I stepped into this research with a background in engineering, but quickly realized that tackling a real-world assistive robotics problem required knowledge across multiple domains—biomechanics, control systems, optimization, simulation, signal processing, and human testing. No single person can be an expert in all of these areas, and this project taught me the true value of collaboration. I am deeply grateful for the guidance of my supervisors and the support of my lab mates, without whom this work would not have been possible.
Throughout the project, I encountered numerous technical and experimental challenges—from unstable system behavior and data inconsistencies to modeling complexities and hardware integration issues. These moments were often frustrating, but they were also where I learned the most. Through each obstacle, I gained a deeper understanding of the research process—not just the technical aspects, but also the perseverance, patience, and creative thinking it demands.
Despite the hardships, the joy of solving difficult problems and developing a technology that can meaningfully improve people’s lives reminded me of why I chose this path. This experience has solidified my passion for assistive technologies and strengthened my resolve to continue this journey in research. As I look ahead to my future as a PhD student, I carry with me not only the skills I’ve gained, but also a renewed sense of purpose—to contribute to meaningful innovations that make a difference in the world.
Conclusion
This project has been a milestone in my academic journey, giving me the opportunity to apply engineering principles to solve real-world challenges in assistive robotics. From developing a personalized human dynamics model to validating its effectiveness through simulation and human testing, the experience has been both technically enriching and personally fulfilling. While the road was far from easy, every setback offered a chance to grow, and every breakthrough reaffirmed my passion for research. I’m excited to carry this momentum forward as I continue exploring the intersection of technology and human well-being in my future endeavors.
Files
The code used for this project can be found in github.