Hansol Kang

Ph.D. Candidate, Mechanical Engineering, Sungkyunkwan University
Robotics Innovatory — advisor: Prof. Hyouk Ryeol Choi

I build learning-based controllers for legged robots. My thesis closes the sim-to-real gap for high-reduction quadrupeds — machines whose gearboxes give them payload and holding torque, but whose actuator dynamics break the assumptions most reinforcement learning pipelines are trained under.

Research

Enhancing Sim-to-Real Transfer for a High-Gear-Ratio Quadruped Robot via Extended Actuator Dynamics Identification

First author · Machines, 2026

Motor modelling alone leaves joint inertia, friction and delay unaccounted for. Here the simulator is fitted to the hardware in a real-sim-real loop: an extended friction model replacing the usual Coulomb-viscous one, with its parameters — per-joint inertia, damping, friction, joint bias — optimized by CMA-ES against measured trajectories.

  • matches real joint motion best exactly where prior models fail, at low joint velocity;
  • on held-out trajectories, joint error drops roughly 33% at HFE and KFE against the previous state of the art;
  • policies trained on the fitted simulator show the smallest sim-to-real gap, with the position-velocity phase portrait deviation cut by up to 42%;
  • combined with a GRU height estimator that infers local terrain from proprioception alone, a policy trained on the fitted simulator clears a 20 cm step and continuous stairs.

Flat terrain

Omnidirectional command tracking on flat ground (2× speed).
Forward velocity command tracking up to 1.1 m/s, with no domain randomization beyond floor friction.
The same controller walking under an 18 kg payload — a third of the robot's own mass.

Rough terrain

Climbing a 20 cm step.
Complex terrain built from stairs and ramps.

Enhancing Sim-to-Real Transfer in High Gear Ratio Quadruped Robot via Short-Term Temporal Information

First author · ICRA 2025, Late Breaking Results

A first pass at the gap, from the actuator inward. I put a BLDC motor model into the simulator instead of an ideal torque source, and gave the policy a time-delayed network over a short window of recent joint states, so it learns the actuator's dynamic response implicitly rather than being told it. The resulting controller tracks user velocity commands on the real AiDIN-VIII (gear ratio > 40:1) and holds up on terrain it never saw flat.

Velocity command tracking on flat ground.
Uneven terrain mocked up with rubber mats (2× speed).

ARS: AI-Driven Recovery Controller for Quadruped Robot Using Single-Network Model

First author · Biomimetics, 2024

Getting a fallen quadruped back on its feet is contact-rich enough that predicting the contact sequence in advance is impractical, so scripted trajectories break down. ARS trains per-behavior policies with a behavior selector, then distills all of them into one network with a DAGGER-like scheme. The single network is cheaper to run online than the multi-network baseline, recovers more smoothly, and generalizes to terrain it never saw in training — 82% success upright, 58% from fully upside-down.

Recovering after being kicked over.
Self-righting from upside-down.

AiDIN quadruped platforms

Robot sensing system and hardware bring-up · co-author, IEEE T-IE 2020, ICRA 2019, IROS 2017–2019

Before the learning work, I took part in developing the AiDIN series as a member of the quadruped robot team at SKKU, including a joint SKKU-UCLA collaboration. My own contribution was the sensing side: designing the ROS-based data acquisition system that brings the IMU, joint encoders and stereo camera into one pipeline for the control and research code to read from — the layer everything else on these robots is built on top of.

I also did the hands-on bring-up — tuning motor drivers, soldering sensing PCBs, assembling the legs, running the experiments.

AiDIN V / VI indoors.
Outdoor trial on campus.

Publications

Journal (SCIE / ESCI)

  1. Hansol Kang, Hyunyong Lee, Jiman Park, Seongwon Nam, Yeongwoo Son, Bumsu Yi, Jaeyoung Oh, Hyeonwoo Yu, Hyouk Ryeol Choi. “Enhancing Sim-to-Real Transfer for a High-Gear-Ratio Quadruped Robot via Extended Actuator Dynamics Identification.” Machines, 2026.
  2. Seongwon Nam, Hyunyong Lee, Hansol Kang, Jiman Park, Yeongwoo Son, Bumsu Yi, Jaeyoung Oh, Hyouk Ryeol Choi. “Real-Time Control-Constrained DDP for Underactuated Balancing of Legged Robots.” IEEE Robotics and Automation Letters, 2026.
  3. Yeongwoo Son, Hyunyong Lee, Hansol Kang, Jiman Park, Seongwon Nam, Jaeyoung Oh, Bumsu Yi, Hyeonwoo Yu, Hyouk Ryeol Choi. “3D RoA-Planner: Path Planner for Quadruped Robots in Confined Spaces Using 3D Rotatable Areas.” IEEE Robotics and Automation Letters, 2026.
  4. Yeong Woo Son, Hyunyong Lee, Hansol Kang, Jiman Park, Seongwon Nam, Jaeyoung Oh, Bumsu Yi, Junha Song, Bogeun Kim, Daegeun Song, Hyouk Ryeol Choi. “RoA-Planner: Rotatable Area based Path Planner in Dense Spaces.” IEEE Transactions on Automation Science and Engineering, 2025.
  5. Soo Yeon Choi, Hyunyong Lee, Ji Man Park, Hansol Kang, Seong Won Nam, Bum Su Yi, Yeong Woo Son, Jae Young Oh, Jun Ha Song, Bo Geun Kim, Dae Geun Song, Hyouk Ryeol Choi. “Development of Two-DOF Simultaneous Gravity Compensator for Mobile Manipulator.” IEEE Robotics and Automation Letters, 2025.
  6. Han Sol Kang, Hyun Yong Lee, Ji Man Park, Seong Won Nam, Yeong Woo Son, Bum Su Yi, Jae Young Oh, Jun Ha Song, Soo Yeon Choi, Bo Geun Kim, Hyun Seok Kim, Hyouk Ryeol Choi. “ARS: AI-Driven Recovery Controller for Quadruped Robot Using Single-Network Model.” Biomimetics, 2024.
  7. Yeong Gwang Son, Tat Hieu Bui, Juyong Hong, Yong Hyeon Kim, Seung Jae Moon, Chun Soo Kim, Issac Rhee, Hansol Kang, Hyouk Ryeol Choi. “CoAS-Net: Context-Aware Suction Network With a Large-Scale Domain Randomized Synthetic Dataset.” IEEE Robotics and Automation Letters, 2023.
  8. Young Hun Lee, Hyunyong Lee, Hansol Kang, Jun Hyuk Lee, Ji Man Park, Chunghyuk Kang, Yoon Haeng Lee, Yong Bum Kim, Hyouk Ryeol Choi. “Balance Recovery Based on Whole-Body Control Using Joint Torque Feedback for Quadrupedal Robots.” Journal of Mechanisms and Robotics, 2021.
  9. Young Hun Lee, Yoon Haeng Lee, Hyunyong Lee, Hansol Kang, Jun Hyuk Lee, Ji Man Park, Yong Bum Kim, Hyungpil Moon, Ja Choon Koo, Hyouk Ryeol Choi. “Whole-body control and angular momentum regulation using torque sensors for quadrupedal robots.” Journal of Intelligent & Robotic Systems, 2021.
  10. Yoon Haeng Lee, Young Hun Lee, Hyunyong Lee, Hansol Kang, Jun Hyuk Lee, Luong Tin Phan, Sungmoon Jin, Yong Bum Kim, Dong-Yeop Seok, Seung Yeon Lee, Hyungpil Moon, Ja Choon Koo, Hyouk Ryeol Choi. “Development of A Quadruped Robot System with Torque-Controllable Modular Actuator Unit.” IEEE Transactions on Industrial Electronics, 2020.
  11. Luong Tin Phan, Yoon Haeng Lee, Young Hun Lee, Hyun-Yong Lee, Hansol Kang, Hyouk Ryeol Choi. “Study on effects of spinal joint for running quadruped robots.” Intelligent Service Robotics, 2020.

Conference

  1. Hansol Kang, Hyunyong Lee, Jiman Park, Seongwon Nam, Yeongwoo Son, Bumsu Yi, Jaeyoung Oh, Bogeun Kim, Daegeun Song, Hyouk Ryeol Choi. “Enhancing Sim-to-Real Transfer in High Gear Ratio Quadruped Robot via Short-Term Temporal Information.” ICRA, Atlanta, 2025 — Late Breaking Results.
  2. Young Hun Lee, Ja Choon Koo, Hyouk Ryeol Choi, Yoon Haeng Lee, Hyunyong Lee, Hansol Kang, Yong Bum Kim, Jun Hyuk Lee, Luong Tin Phan, Sung Moon Jin, Hyungpil Moon. “Whole-Body Motion and Landing Force Control for Quadrupedal Stair Climbing.” IROS, 2019.
  3. Seungjun Woo, Jinjae Shin, Yoon Haeng Lee, Young Hun Lee, Hyun-Yong Lee, Hansol Kang, Hyouk Ryeol Choi, Hyungpil Moon. “Stair-mapping with Point-cloud Data and Stair-modeling for Quadruped Robot.” UR, 2019.
  4. Yoon Haeng Lee, Young Hun Lee, Hyun-Yong Lee, Hansol Kang, Luong Tin Phan, Sungmoon Jin, Yong Bum Kim, Dong-Yeop Seok, Seung Yeon Lee, Hyungpil Moon, Ja Choon Koo, Hyouk Ryeol Choi. “Force-controllable Quadruped Robot System with Capacitive-type Joint Torque Sensor.” ICRA, 2019.
  5. Luong Tin Phan, Yoon Haeng Lee, Young Hun Lee, Hyun-Yong Lee, Hansol Kang, Hyouk Ryeol Choi. “Effects of spinal joint on quadrupedal bounding.” URAI, 2018.
  6. Luong Tin Phan, Yoon Haeng Lee, Young Hun Lee, Hyunyong Lee, Hansol Kang, Hyouk Ryeol Choi. “Study on quadruped bounding with a passive compliant spine.” IROS, 2017.
  7. Young Hun Lee, Yoon Haeng Lee, Hyunyong Lee, Luong Tin Phan, Hansol Kang, Uikyum Kim, Jeongmin Jeon, Hyouk Ryeol Choi. “Trajectory design and control of quadruped robot for trotting over obstacles.” IROS, 2017.
  8. Yoon Haeng Lee, Young Hun Lee, Hyunyong Lee, Luong Tin Phan, Hansol Kang, Yong Bum Kim, Hyouk Ryeol Choi. “Development of torque controllable leg for running robot, AiDIN-IV.” IROS, 2017.

Awards

Projects

Education