|
Deep Reinforcement Learning based Autonomous Air-to-Air Combat using Target Trajectory Prediction |
|
Jaewoong Yoo(KAIST, Korea), Donghwi Kim, Jedsadakorn Yonchorhor, David Hyunchul Shim(Korea Advanced Institute of Science and Technology, Korea) |
| The purpose of this study is to design an intelligent control system based on reinforcement learning for autonomous air-to-air combat. In this work, Long short-term memory (LSTM) was applied to Proximal Policy Optimization (PPO) algorithm in order to refer not only the current state of the target but also the previous ones. In addition, target trajectory prediction was performed to occupy an advantageous location from enemy aircraft in close engagements. The result of the study confirmed that the maneuvers of trained agent were similar to the performance of human pilots in Digital Combat Simulator (DCS). |
|
|
Performance Analysis of Momentum-based Optimizers for Face Recognition using Deep Learning |
|
Hanul Son, Seungbin MOON(Sejong University, Korea) |
| The recent COVID-19 pandemic has enforced mass usage of facial masks in public settings. This has lead to a degradation in performance of commercial face recognition systems such as access control systems, smartphone recognition, mobile robots etc. Removing the mask for face registration and verification is not only inconvenient but poses a health risk to the users too. Robust face recognition systems need to be optimized for normal and masked face images. In this paper, we explore the effects of using different optimizers on the performance of deep convolutional neural networks. We compare the converging time and final accuracy of momentum based and step size based optimizers. |
|
|
Development Of The Orthodontic Wire Processing Robot : From CT Image To Wire Processing |
|
Jongchan Park, Nchumpeni Chonpemo Shitiri, Huijun Jeon, Youngwoo Kim(Korea National University of Transportation, Korea) |
| We proposes the development of the software for orthodontic professionals that aid the design of the orthodontic wire shapes, a robot that produces the shape of the wire that has already been designed by the orthodontist, and a simulation/ control platform for orthodontic wire processing robot has been described in this paper. Furthermore, we have developed a mathematical model to obtain the wire’s machining point setting, bending direction, and bending angle using a homogeneous transformation functions. Based on the processed data, the wire is then fed into the robot, which performs the wire bending operations. |
|
|
Central Pattern Generator Optimization using Gradient Descent Method |
|
Karthik Murugesh, Youngwoo Kim(Korea National University of Transportation, Korea) |
| In this paper, we discuss the process of parameter optimization for various Gait patterns using the gradient descent technique, the Sensory Feedback System in the neuro-musculoskeletal model can simulate somatosensory input such as angle and angular displacement for individual joints and links. Using a neural model structured with a number of parameters we employ the gradient decent approach in the optimization process. In gait optimization for biped robots and wearable exoskeletons, a neural model based on a deep learning approach is used to optimize gait pattern for robots in numerous applications and various external ground contact conditions. |
|
|
Modeling of Adaptation Mechanism for Customized Lower Limb Rehabilitation Platform |
|
Shamanth S, Youngwoo Kim(Korea National University of Transportation, Korea) |
| We demonstrate a customizable lower limb rehabilitation platform that improves the efficacy of rehabilitation training. The two-stage algorithm is used to obtain optimized trajectory, which is crucial in establishing the robotic interaction with the patient's range of motion capabilities. We demonstrate a metaheuristic approach utilizing a genetic algorithm to generate optimal solutions for the robot's interaction with its external environment. The developed simulation environment, as well as its configurable musculoskeletal model and gait trajectory analysis and optimization tools, which are tailored to each user's needs are also presented. |
|
|
GTOE : Anomaly Detection Geometric Transformation Network with Outlier Exposure for Real Industry Data |
|
Yong Wan Kwon, Dong Joong Kang(Pusan National University, Korea) |
|