Jeonghyeon Wang

Study of general local path planner to overcome U-shaped enclosure problem

A new real-time path planning method for mobile robot is proposed based on weighted virtual tangential vector (WVTV) algorithm. The WVTV is obtained by computing weighted sum of virtual tangential vector (VTV) from detected obstacles. It has the property of keeping driving safety and achieve multi object optimization problem with linear combination of several vectors. In spite of these advantages, it is not much general solution to apply any of the mobile robot system. Also, it cannot solve re-entrance of u-shape problem. Our suggested algorithm is considered these pros and cons of WVTV and analyzing the property of obstacle shape, we will have more possibility to solve u-shape problem. Once we classify obstacle as a convex and non-convex, expected route of mobile robot will be identified. Significant performance enhancement of our algorithm is demonstrated by using Matlab simulator modeled by Khepera III for any complex obstacles and mixing any applied u-shape obstacles.

Development of an Unmanned Surface Vehicle System for the 2014 Maritime RobotX Challenge

This paper addresses the development of an unmanned surface vehicle (USV) system by Team Angry-Nerds from KAIST for the inaugural Maritime RobotX Challenge competition, which was held on October 20-26, 2014, in Marina Bay, Singapore. The USV hardware was developed on a catamaran platform by integrating various system components, including propulsion, sensors, computer, power, and emergency systems. The competition comprised five mission tasks: 1) navigation and control, 2) underwater search and report, 3) automatic docking, 4) buoy search and observation, and 5) obstacle detection and avoidance. Onboard intelligence was a key factor for all of the mission tasks which needed to be performed autonomously with no human intervention. Software algorithms for vehicle autonomy were developed, and executable computer codes were implemented and integrated with the developed USV hardware system. This paper describes the development process of the USV system and its application to the competition mission tasks.

Semantic segmentation of urban scenes with enhanced spatial contexts

This paper proposes a method to solve the image segmentation problem effectively by applying local spatial relations. The proposed method is based on the conjecture that spatially neighboring objects have a higher probability of relating each other than nonadjacent objects. Using depth information from 3D lidar, spatial contextual information between image segments is obtained. This information is modeled in the framework of Conditional Random Field to penalize distant objects that are likely to have no particular relation between them. The proposed method is evaluated on the publicly available image dataset, and it shows better global and class classification rates compared to the previous state-of-the-art work.

Three-dimensional reconstruction of a semi-submersible offshore platform with an unmanned surface vehicle

This paper addresses the three-dimensional (3D) reconstruction of a floating structure with an unmanned surface vehicle (USV). Onboard lidar and sonar sensors are employed to collect a volumetric point cloud of the structure both above and below the waterline. These measurements are obtained in the vehicle-fixed frame; thus, for successful 3D reconstruction, precision estimation of trajectory and attitude is required. GPS signals are severely deteriorated or unavailable near and under floating structures. Therefore, relative navigation with respect to the planar surfaces of their hull structures is performed in the framework of simultaneous localization and mapping (SLAM). This approach enables high-precision navigation and mapping near and under a large floating structure. A field experiment was performed in a semi-submersible offshore platform environment and the results are presented.

Optimization of fish-like locomotion using hierarchical reinforcement learning

With an interest in advanced marine propulsion systems, much research has been done on mimicking fish-like locomotion using flapping fins. This study aims to optimize the swimming pattern of fish-like locomotion based on hierarchical reinforcement learning. A simplified carangiform fish model is employed and a segmented tail motion is learned by Q-learning to maximize the average forward velocity by flapping the tail fin. The performance of the self-learned swimming pattern is verified and analyzed in terms of the flapping efficiency. The results show that the flapping angle limit of approximately 35 degrees is best in maximizing the forward moving velocity and the hierarchical reinforcement learning approach is effective in providing a reasonable solution for a large-scale problem.

Scalar field reconstruction based on the Gaussian process and adaptive sampling

This paper proposes a method of reconstructing a scalar field by adaptively choosing sampling locations and using the measurements obtained from those locations to reconstruct an estimate of the underlying field using Gaussian process regression. Spreading sampling points evenly over the field may not always be effective if the field is not uniformly distributed and the maximum number of measurements is limited. Taking more measurements in regions of large changes in the field than in regions of small changes can give a better estimate than spreading the same number of measurements evenly over the space. The proposed algorithm was tested on a synthetic scalar field and compared to two popular methods of determining sensor placement based on entropy and mutual information from information theory.

Constrained motion planning for robot manipulators using local geometric information

This paper proposes a new motion planning algorithm for robot manipulator systems with path constraints. The constraint function of a manipulator determines the subspace of its joint space, and a proposed sampling-based algorithm can find a path that connects valid samples in the subspace. These valid samples can be obtained by projecting the samples onto the subspace defined by the constraint function. However, these iteratively generated samples easily fall into local optima, which degrades the search performance. The proposed algorithm uses the local geometric information and expands the search tree adaptively to avoid the local convergence problem. It increases the greediness of the search tree when it expands toward an unexplored area, which produces the benefit of reducing computational time. In order to demonstrate the performance of the algorithm, it is applied to two example problems: a maze problem using PUMA 560 under predefined constraints and a closed-chain problem using two Selective Compliance Assembly Robot Arms. The results are compared with those obtained with an existing algorithm to show the improvement in performance. Graphical Abstract

A RRT-based collision-free and occlusion-free path planning method for a 7DOF manipulator

In this paper, we present a path planning algorithm of a 7DOF manipulator with RRT(Rapidly-exploring Random Trees). The method presents a occlusion-free and collision-free path planning algorithm. First, we present a basics of RRT algorithm which is a basic searching method for a path planning of this paper. Second, the implementation of a occlusion-free path constraints is presented. Third, the collision-avoiding path planning constraints is presented. Fourth, the integration of RRT with these two constraints is shown. The proposed method is verified by a numerical simulation with Matlab and v-rep robotics simulation tool.