Donghyeon Kim

Donghyeon Kim

Auto-generated profile, based on publicly available information

Donghyeon Kim

Department of Intelligence and Information, Graduate School of Convergence Science and Technology, Seoul National University, Suwon-si, Republic of Korea

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
21 December 2021 in IEEE Transactions on Robotics
Verified
Journal Article
Closed access
DKDonghyeon Kim
Donghyeon Kim
DLDaegyu Lim
Daegyu Lim
JPJaeheung Park
Jaeheung Park
As human-robot collaboration increases and robots are applied to everyday life, interest in safety issues is increasing. To be safely used in real life, in addition to collision prevention algorithms, robots need to quickly detect unexpected collisions and take appropriate actions. Recently, deep learning-based collision detection algorithms have been proposed to overcome the limitations of model-based collision detection methods, but there are also limitations on deep learning methods, especially in data collection. The collected data are often insufficient because collecting collision data is laborious and intrinsically imbalanced, meaning that the collision data are much smaller than the free-motion data. Moreover, since collecting collision data is risky and might cause potential damage to the robot, applying the deep learning method to a new target robot on mass production is highly restricted. Therefore, in this article, an inductive bias is imposed on network structure and input variable to be sample-efficiently trained, which is suitable for insufficient imbalanced data. The proposed modularized neural network removes the connection between other joints at the front part of the network, reducing the search space of the learnable parameter. Moreover, the input variable is selected to take both dynamics features and error-related features into account. Consequently, the proposed method is versatile, showing successful generalization performance to random motion, random collision location, and various loads at the end-effector. Furthermore, to circumvent the limitations of applying deep learning methods to mass production, transfer learning is proposed, which does not require any collision data from the target robot. The proposed data-mixture method and collision ratio adjustment method for fine-tuning are validated with two source robots and one target robot. The transferred network is effective to be applied on mass production without losing performance compared to a specific-robot-trained network(a network trained with specific robot data and applied to the same robot).

11 August 2023 in IEEE Robotics and Automation Letters
Verified
Journal Article
Closed access
DKDonghyeon Kim
Donghyeon Kim
GBGlen Berseth
Glen Berseth
MSMathew Schwartz
Mathew Schwartz
JPJaeheung Park
Jaeheung Park
In this letter, we review the question of which action space is best suited for controlling a real biped robot in combination with Sim2Real training. Position control has been popular as it has been shown to be more sample efficient and intuitive to combine with other planning algorithms. However, for position control, gain tuning is required to achieve the best possible policy performance. We show that, instead, using a torque-based action space enables task-and-robot agnostic learning with less parameter tuning and mitigates the sim-to-reality gap by taking advantage of torque control's inherent compliance. Also, we accelerate the torque-based-policy training process by pre-training the policy to remain upright by compensating for gravity. The letter showcases the first successful sim-to-real transfer of a torque-based deep reinforcement learning policy on a real human-sized biped robot.

12 December 2024 in IEEE Robotics & Automation Magazine
Verified
Journal Article
Closed access
DKDonghyeon Kim
Donghyeon Kim
HLHokyun Lee
Hokyun Lee
JCJunhyeok Cha
Junhyeok Cha
JPJaeheung Park
Jaeheung Park
Reinforcement learning (RL) offers a promising solution for controlling humanoid robots, particularly for bipedal locomotion, by learning adaptive and flexible control strategies. However, direct RL application is hindered by time-consuming trial-and-error processes, necessitating training in simulation before real-world transfer. This introduces a reality gap that degrades performance. Although various methods have been proposed for sim-to-real transfer, they have not been validated on a consistent hardware platform making it difficult to determine which components are key to overcoming the reality gap. In contrast, we systematically evaluate techniques to enhance RL policy robustness during sim-to-real transfer by controlling variables and comparing them on a single robot to isolate and analyze the impact of each technique. These techniques include dynamics randomization, state history usage, noise/bias/delay modeling, state selection, perturbations, and network size.We quantitatively assess the reality gap by simulating diverse conditions and conducting experiments on real hardware. Our findings provide insights into bridging the reality gap, advancing robust RL-trained humanoid robots for real-world applications.

28 June 2024 in International Journal of Social Robotics
Verified
Journal Article
Open access
PDF file
BPBeomyeong Park
Beomyeong Park
DKDonghyeon Kim
Donghyeon Kim
DLDaegyu Lim
Daegyu Lim
SPSuhan Park
Suhan Park
JAJunewhee Ahn
Junewhee Ahn
JPJaeheung Park
Jaeheung Park
Avatar robots enable the teleoperation and telepresence of an operator with a rich and meaningful sense of existence in another location. Robotic avatar systems rely on intuitive interactions to afford operators comfortable and accurate robot control to perform various tasks. The ability of operators to feel immersed within a robot has drawn interest in multiple research fields to explore the future capabilities of such systems. This paper presents a robotic avatar system based on a custom humanoid robot, TOCABI, with a mobile base. Its teleoperation system was developed in response to the ANA Avatar XPRIZE. Combining the life-size humanoid robot and the mobile base allows for improved mobility and dexterous manipulation. The robotic avatar system comprises the robot/base and an operator station that incorporates haptic feedback devices, trackers, a head-mounted display, gloves, and pedals. These devices connect the robot-environment interaction and operator-avatar robot experience through visual, auditory, tactile, haptic, and kinesthetic feedback. Combining the untethered battery-operated and Wi-Fi-enabled robot with these sensory experiences enables intuitive control through the operator’s body movement. The performance of the robotic avatar system was evaluated through user studies and demonstrated in the ANA Avatar XPRIZE Finals, represented by Team SNU, where it completed 8 of the 10 missions, placing the team eighth among the 17 finalists.