Hanjun Kim

Monitoring human status using sensor fusion and AI techniques in real-world scenarios

2022–present · Human Biomechanics · Robot Control · AI

Defining a "good" wearable robot often depends on how it impacts the human user. Human performance can be assessed through metrics such as energy expenditure, biological joint work, and walking speed. Accordingly, many researchers tune robot parameters based on these outcomes, which are typically measured in controlled laboratory settings. However, the lack of such measurements presents a major challenge for outdoor applications. To address this, we have integrated multiple sensors to capture users' movements and developed AI frameworks that estimate key human performance outcomes in real-world environments. These estimations not only enable evaluation of current performance but also allow controllers to be automatically adapted, since changes in human dynamics require different levels of assistance.

Publications: IEEE RAL (slope estimation), IEEE TMRB (parameter estimation), IEEE TMRB (speed estimation)

Conference: 2024 ASB Abstract (slope estimation)

A Continuous, Biomimetic Impedance Control Strategy for Robotic Leg Prostheses

2023–present · Human Biomechanics · Robot Control

Traditionally, prosthesis controllers were designed around separate walking tasks for different terrain modes, such as level ground, ramps, and stairs. These devices detected the user's environment and switched to the corresponding controller. However, human lower-limb dynamics vary not only across modes but also within them (e.g., walking on a 5° ramp vs. a 13° ramp). As a result, this mode-based strategy often led to errors in joint angles and torques. To overcome this limitation, we developed a unified approach that represents all modes within a continuous slope domain, smoothly connecting human dynamics across terrains. This continuous biomechanics framework allows prostheses to derive biomimetic, adaptive control parameters.

Publication: Clinical Biomechanics (controller parameterizing)

Conference: ICORR 2025 (human-inspired controller design)

Machine Anomaly Detection Using Sound Spectrogram Images and Neural Networks (Master's Thesis Topic @Purdue)

2017–2019 · Signal Processing · AI

Unexpected machine failures on the shop floor can severely reduce productivity and, in worse cases, cause dangerous accidents. To prevent such incidents, factories often rely on mechanical sensors to monitor machine health. However, commercial solutions are typically expensive and demand additional conditioning costs. Moreover, effective physical models or AI frameworks require large datasets of faulty machine behavior, which are difficult to collect in practice. To address these limitations, we developed a cost-effective sensing method using commercial stethoscopes and an autoencoder-based AI framework that does not require large amounts of fault data.

Publications: *MS Thesis, Journal of Intelligent Manufacturing, Proceeding from NAMRC 48

Patent: US11475910B2

System Integration for Multi-sensor Fusion using ROS2 Framework

Representing human movement in outdoor environments requires integrating multiple sensors and centralizing them for real-time interaction between the human and the wearable robot. We are developing a modular ROS2 framework with intact-side and prosthetic-side systems, along with a sync module for in-lab motion capture (MoCap). The Jetson Orin Nano handles high-level integration and onboard computation, while the intact side and MoCap modules focus on signal acquisition and transfer.

Human Gait Event Tracking with IMU-Driven Relative Leg Geometry

Human leg dynamics vary throughout the gait cycle, making real-time detection of events such as heel contact, toe-off, and single support essential for adaptive control of wearable devices, particularly prostheses. A challenge is that gait environments use different criteria for defining these events-for example, knee angle on stairs versus ankle angle on ramps. To overcome this, our centralized system leverages the relative geometry between the intact and prosthetic sides, which exhibits consistent cyclic patterns across environments. This bilateral approach enables a unified, environment-independent method for gait event detection.

Hardware enhancements for Open Source Leg (OSL)

To leverage publicly available resources, we have adopted and modified the Open Source Leg (OSL), originally developed at the University of Michigan. Our enhancements focus on improving machinability, user-friendliness, and power source integration.