01
Exoskeleton system design · prototyping
6-DoF articulated arm (2 clavicle + 3 shoulder + 1 elbow); custom actuator design integrated with off-the-shelf robotics components from Maxon, Welcon Systems, Robotous, and others.
Minwoo Lee · Robotics Control Engineer · April 2026
Hi, I’m Minwoo — I build
contact-rich robots end-to-end.
From mechanical design and haptics to 800 Hz real-time force control,
I focus on making physical interaction controllable for manipulation.
I focus on making physical interaction controllable for manipulation.
Minwoo Lee / Robotics Control Engineer
Summary • Edu & Career • Agenda
Day1-Ready Engineer for Robot Control and Systems Integration + Fast-Learner grounded in Advanced Dynamics and Embodied AI
Education & Career
- M.S.
Kyung Hee Univ 2024.03 – 2026.02 - M.S. in Software Convergence | Advisor : Prof. Seungjae Oh
(Thesis) Building Blocks for Full Upper-Limb Haptics —
integrated study of an upper-limb exoskeleton + a pneumatic haptic glove - Founded 'Realimerse'
2022 – 2024.02 - Co-founder / CTO —
Led a 4-person mechanical · electrical · control team.
Live wearable-robot prototype demo at CES 2024. - B.S.
Kyung Hee Univ., 2016.03 – 2024.02 - B.S. in Mechanical Engineering + Software Convergence
Mechanical dynamics · control + a robot-mobility track in Software Convergence —
A robotics-focused CSE curriculum: Linux · ROS · CV · ML · RL. - Publications · Patents
- SCIE registered journal (Co-author) — Adv. Mater. Technol. '23
International conference demonstration — IEEE ISMAR '25
Domestic undergraduate paper award — ICROS '22
A granted patent ('25.09 metaverse HMI robotic device)
A Filed patent ('26.01 bidirectional pneumatic glove)
Portfolio Agenda
- A · Proof —
Exoskeleton - Force-controlled exoskeleton robot — 6-DoF task-space force control, 800 Hz PREEMPT_RT real-time control, CANopen-based HAL
- B · Proof —
Glove - Closed-loop pneumatic glove — ±100 kPa bidirectional, 173 ms closed-loop, integrated intelligent interaction, N=16 user study
- C · Synthesis
- Sensing-rich HMI thesis — A human-machine interface that fuses the exoskeleton and the glove to collect the physical dataset needed for force-aware, contact-rich robot policy learning.
- D · Plus
- AX-driven productivity — Knowledge Lake · personal LLMOps infra · agentic development harness
Page 02 / 9
Minwoo Lee / Robotics Control Engineer
A PROOF · Force-Controlled Exoskeleton
An upper-limb exoskeleton for task-space force control — Impedance Display: full-stack implementation and applications
2024.03 – 2026.02
Scope of work — full-stack implementation across Mechanical Design · HAL Implementation · RT Control · Force Control Task · Simulation & Validation
02
Hardware abstraction layer
High-rate CAN/CANopen comms; ros2_control C++ HAL; drivers built on the hardware_interface base class.
03
Real-time control implementation
Deterministic 800 Hz real-time stack on PREEMPT_RT with SCHED_FIFO multithreading.
04
Task-space force control
Lie-group dynamics on Pinocchio C++ RBD; gravity and dynamics compensation closed by a 6-axis F/T sensor.
05
Controller validation in Gazebo simulation
01Wearing demo — upper-limb exoskeleton
02Lie-group POE kinematics · human-alignment optimization
03Gazebo simulation validation
04Redundancy-resolution control for multi-contact rendering in task space
Page 03 / 9
Minwoo Lee / Robotics Control Engineer
A PROOF · Exoskeleton System Stack
Multi-DoF · Real-time Linux · Task-Space Force Control — Hands-on experience across the full system
ros2_control HAL · PREEMPT_RT · 1 kHz CANopen · Pinocchio RBD — a 4-layer stack I designed, implemented, and validated end-to-end
L0 · Hardware
Physical Hardware
- Actuator Modules
- Motor Controllers
- F/T Sensors
SocketCAN
read / write
L1 · HAL
Hardware Abstraction
- ROS2 hardware_interface
- CAN / CANopen
- Read / Write Handlers
- RT Deadline Logic · faults
states / commands
L2 · Control
Controller Implementations
- Dynamics compensation
(gravity, Coriolis) - Force rendering
- Pinocchio RBD library
topics · params
L3 · Application
ROS Application
- robot_description + tf2
- Topic interfaces
- Diagnostics
- Application integration
DETERMINISTIC EXECUTION
RT Linux · Scheduling
PREEMPT_RT kernel
SCHED_FIFO priorities
CPU affinity · IRQ tuning
memlock · rtprio limits
Control rate
800Hz
Deterministic loop on PREEMPT_RT + SCHED_FIFO
CAN sync
1000µs
3 independent CAN channels · CANopen CiA 402 · TPDO/RPDO
Peak / continuous torque
9 / 3 Nm
Cartesian force rendering at the 1 kgf class
System mass
2.85kg
Aluminum frame + double-parallelogram RCM linkage
Hardware
CubeMars AK60-6 (shoulder, clavicle) · maxon ECX FLAT 32L + HD CSF-11 50:1 (elbow). 3D-printed + aluminum frame; double-parallelogram RCM linkage aligns the shoulder ball joint to a remote center of rotation.
Control SW
PREEMPT_RT Linux + SCHED_FIFO · ROS2 ros2_control + custom hardware_interface (C++17) HAL · 3 independent CAN channels @ 1 kHz (CANopen CiA 402)
Dynamics
Pinocchio RBD · SE(3) POE · variable-weighted Jacobian for tunable per-joint involvement · RNEA · gravity compensation
Application control
Human/environment task-space contact-force control — Active transparency control + task-space wrench reference tracking. Closed-loop using a Robotous RFT64 6-axis F/T sensor.
Page 04 / 9
Minwoo Lee / Robotics Control Engineer
B PROOF · Hand-Pose Sync Glove
Leader-follower hand-pose sync for skill transfer — a closed-loop pneumatic-glove system and its control methodology
2024.09 – 2025.12
Scope of work — integrated implementation of design · closed-loop sync · AI scenarios · user-study validation
01
Pneumatic-glove system design · prototyping
4-channel bidirectional pneumatic glove (thumb · index · middle · ring+pinky) + Manus Quantum 15-DoF hand-pose feedback. Lightweight pneumatic system integrating positive/negative buffer tanks with a single compressor.
02
Closed-loop follower synchronization
A ±100 kPa bidirectional pneumatic glove tracks the master hand pose within 173 ms under PI control.
03
Advanced scenarios
Intelligent interaction built on VLM-based user-intent inference and a hand–object grasp-synthesis pipeline.
04
Skill-transfer validation
IRB-approved N=16 study. Statistical validation of novelty, agency, and AI-guidance.
01Closed-loop hand-pose synchronization: master pose → PI control → pneumatic command → follower pose
02Glove interface with pneumatic actuators + supporting hardware
03Positive/negative buffer tanks + single-compressor pneumatic system (Festo VEAB + DAC)
04Sensing — hand-pose motion capture (Manus Quantum), ArUco
Page 05 / 9
Minwoo Lee / Robotics Control Engineer
B PROOF · Glove Control · Experiment
From hand-pose synchronization to intelligent hand–object interaction
Closed-loop synchronization combined with vision-language intent inference + hand–object grasp synthesis — skill-transfer effectiveness validated in an N=16 user study
Synchronization in action · master → follower
Sequence: an expert hand pose reproduced by the bidirectional pneumatic glove
Sub-interaction · AI scene understanding integration
Scenario variety — segmentation · 3D reconstruction · grasp synthesis
End-to-end latency
173ms
Target pose → settled within ±5%
Pressure range
±100kPa
Bidirectional — pressure and vacuum
Pneumatic system form factor
4.53kg
400×200×150 mm · −45% volume / −30% mass vs. prior work
Channel independence
4 × independent
Festo VEAB regulators on thumb · index · middle · ring+pinky
Advanced scenarios · scenario expansion
1. Egocentric vision + eye tracker for gaze and environmental context.
2. VLM-based high-context inference of user intent from hand pose — gesture-driven binary responses, sentiment from compliance with imposed hand poses, and more.
3. Augmented hand–object interaction — image segmentation + RGB2Mesh reconstruction of nearby objects, plus an optimal grasp-pose synthesis pipeline.
Validate · skill-transfer evaluation
IRB-approved N=16 user study. Core values of the skill-transfer method — novelty d=1.12 (large, p<.001) and guidance intuitiveness d=0.97 (large, p=.002), both verified at a large effect size.
Sense of agency d=−0.63 to −1.32 (the AI-reliance trade-off detected simultaneously).
Page 06 / 9
Minwoo Lee / Robotics Control Engineer
C Synthesis of the two preceding studies
A sensing-data-rich HMI proposal for imitation learning
Exoskeleton (arm) + glove (dexterity) — adding force and tactile to motion: a whole-upper-limb bilateral human-machine interface
01
Traditional teleopStructural limits of the conventional approach
Unilateral, motion-level
• No remote-env reflection; contact sense is severed
End-effector only, heterogeneous DoF
• Human biomechanics is lost
Stationary + reduced agency
• Unnatural posture, accumulated fatigue
• No remote-env reflection; contact sense is severed
End-effector only, heterogeneous DoF
• Human biomechanics is lost
Stationary + reduced agency
• Unnatural posture, accumulated fatigue
VR HMD operator
Pedestal leader-arm
Kinesthetic teaching
02 — Ours
Exoskeleton + glove
Acquired data streams
• Kinematics & force/wrench: high-rate, HW-anchored
• Dexterity: high-DoF hand pose + tactile
• Exoskeleton as a sensing platform
Differentiators
• As-is engagement — no HMD, natural posture
• Bilateral — remote wrench reflected
• Whole-arm coverage — shoulder to fingertip
• Sense of agency preserved
• Kinematics & force/wrench: high-rate, HW-anchored
• Dexterity: high-DoF hand pose + tactile
• Exoskeleton as a sensing platform
Differentiators
• As-is engagement — no HMD, natural posture
• Bilateral — remote wrench reflected
• Whole-arm coverage — shoulder to fingertip
• Sense of agency preserved
force
Exoskeleton
+
dexterity
Glove
From Quality,
To Scalability
To Scalability
1. Beyond the motion-only paradigm
VLA and other motion-based policies are scaling to internet-scale, yet the data still lives at the motion level — failing to capture the contact states, resistance, friction, and fine compliance strategies that arise in assembly, insertion, and fastening.
The exoskeleton + glove add force, wrench, and tactile channels → the sim2real gap closes under force-rich supervision.
2. Surrogation of manual labor
Physical strain on the floor vs. the friction of teleoperation — experts need a reason to use the interface. With feedback-rich, ergonomic design, once operating ease crosses the threshold the shift to remote work follows naturally.
3. Mass-scale data acquisition
Teleoperation becomes the medium through which human skill accumulates — assembly, precision manufacturing, medicine, and other domain expertise gets corpusified automatically.
VLA and other motion-based policies are scaling to internet-scale, yet the data still lives at the motion level — failing to capture the contact states, resistance, friction, and fine compliance strategies that arise in assembly, insertion, and fastening.
The exoskeleton + glove add force, wrench, and tactile channels → the sim2real gap closes under force-rich supervision.
2. Surrogation of manual labor
Physical strain on the floor vs. the friction of teleoperation — experts need a reason to use the interface. With feedback-rich, ergonomic design, once operating ease crosses the threshold the shift to remote work follows naturally.
3. Mass-scale data acquisition
Teleoperation becomes the medium through which human skill accumulates — assembly, precision manufacturing, medicine, and other domain expertise gets corpusified automatically.
Page 07 / 9
Minwoo Lee / Robotics Control EngineerD · AX
D AX-driven productivity
Beyond the capacity of a single engineer — an AX-driven development workflow
The bottleneck is the distinctly human work of holding a vision, imagining it, and moving it into execution
01
Knowledge Lake: turning tacit knowledge into data
I embed my coursework notes, daily journals, and research references as semantic vectors, then cluster them under an ontology — turning years of personal material into a transparently queryable knowledge source.
02
Personal Agent Infrastructure
I run OpenClaw on a personal home-lab server (RPi), backed by a cloud VPS, VPN, and supporting network infrastructure — a fully private agent that keeps work continuous across environments. Building on this hands-on understanding of agentic workflows, I am broadening into data governance, security, and wearable/IoT hyper-connectivity, aiming at fully personalized technology.
03
Harnessing Agentic Power for Development
What used to take all-nighters to translate from idea to software now runs through coding agents like Claude Code and Codex. I go past basic CLI use — orchestrating parallelism, branching, and tool use into an agentic harness that delivers code and documents end-to-end, turning my SW capability directly into output.
VISUALIZATIONKnowledge Lake — 2D UMAP projection · ontology-based classification overlay
Page 08 / 9
Closing anchor
Across two axes of work — exoskeleton and glove — what I've come to see is one thing: Force-Augmented Human Demonstration, conducted through a robotic medium.
Internet-scale data has carried LLMs to where they are.
The same playbook will not carry VLA the rest of the way
— Demonstration quality will.
And quality, for contact-rich tasks, means more than motion.
Capturing the fine manipulation of the hand, the forces along the arm, the tactile signature of contact — turning all of it into data — is the prerequisite
for any policy that has to actually touch the world.
I want to design and build the means to capture that data, refine the methods that learn from it, and — on the stage that ultimately brings robotics into the mainstream — push contact control and imitation learning all the way through.
Minwoo Lee
Robotics Control Engineer
[email protected]
mechaminu.dev
mechaminu.dev
Page 09 / 9