|
Semantic Descriptors into Representation for Robust Indoor Visual Place Recognition |
|
Nuri Kim, Minjae Kang, Songhwai Oh(Seoul National University, Korea) |
| Visual place localization (VPL) is a problem finding the closest database image from a query image. Since the outdoor images can be recognized from GPS sensors, VPL in an indoor scene is a difficult problem. Also, Image changes indoors are more severe than outdoors. It is because the position of objects can be easily changed indoors. To tackle this problem, we propose a novel localization dataset with 3D objects considering their physical locations in a scene and encode semantic information using neural networks. Experimental results show that our proposed method outperforms other baseline methods on our localization dataset. |
|
|
Exploring Advanced Process Equipment Visualization as a Step Towards Digital Twins Development: A CFD-DNN Approach |
|
Dela Quarme Gbadago, Jiyoung Moon(Inha University, Korea), Sungwon Hwang(INHA University, Korea) |
| Equipment performance monitoring still faces difficult challenges due to the limitations of sensors and modelling techniques. Hence, we presented an efficient deep neural network-based 3D visualization platform analogous to the digital for detailed process equipment visualization in real-time. The methodology included rigorous physics-based mathematically modelling in the form of computational fluid dynamics simulations and artificial neural network surrogate models. A visualization platform was created in the form of a graphical interface using python. |
|
|
The use of Neural Network for Nonlinear Predictive Control design for and Overhead Crane |
|
Jakub Nemcik(Technical University of Ostrava, Czech Republic), Filip Krupa(VSB-TU Ostrava, Czech Republic), Stepan Ozana, Ivan Zelinka(Technical University of Ostrava, Czech Republic), Zdenek Slanina(VSB-TU Ostrava, Czech Republic) |
| The importance of nonlinear model predictive control (NMPC) implementations for industrial processes rises with the increasing of computational power in all hardware units used for regulation and control in practice. However, it assumes a sufficiently accurate model. In case of more complex systems, there might be problem to perform analytical identification. Instead of this, numerical approaches may be deployed with benefit. This paper deals with the design of NMPC for a nonlinear model of an overhead crane using a neural network and compares the solution with the one achieved with the use analytical model of the system. |
|
|
Discrete Task-Space Automatic Curriculum Learning for Robotic Grasping |
|
ANIL KURKCU, CIHAN ACAR(I2R, Singapore), DOMENICO CAMPOLO(NTU, Singapore), KENG PENG TEE(I2R, Singapore) |
| We present an automatic curriculum learning algorithm for discrete task-space scenarios. Our curriculum generation is based on difficulty measure between tasks and learning progress metric within a task. We apply our algorithm to a grasp learning problem involving 49 diverse objects. Our results show that a policy trained based on a curriculum is both sample efficient compared to learning from scratch and able to learn tasks that the latter could not learn within a reasonable amount of time. |
|
|
Determination of Defects for Dynamic Objects Using Instance Segmentation |
|
Seongho Jin, Jangmyung Lee(Pusan National University, Korea) |
| In this paper, object recognition/analysis using artificial intelligence in the smart factory is used for real-time quality inspection in the manufacturing process. Determining defects in the manufacturing process is a very important process. We use deep learning-based CNNs to build detection models for dynamic defect detection. To track the dynamic object placed on the conveyor, the camera is fixed to the object to be identified, and at the same time, the state of the object learned through deep learning is distinguished to determine the defect. Accordingly, a system for real-time tracking and defect determination on the analyzed defective quality goods is established. |
|
|
A Deep Reinforcement Learning Algorithm based on Modified Twin Delay DDPG Method for Robotic Applications |
|
Carlos Alberto Vasquez Jalpa, Mariko Nakano, Enrique Escamilla-Hernandez(Instituto Politecnico Nacional, Mexico) |
| This research is focused on training an agent, with Deep Reinforcement Learning (DRL) whose objective is to find a certain object in an environment, to achieve this the following stages were carried out: exploration stage, where a proximity sensor is used that helps to collect data from the environment, visual learning stage, in which a CNN is used in charge of classifying the images obtained, training stage, which uses a block of some variables related to the agent and environment, using Deep Q-Learning, DRL and Gradient Policy, and inference stage. Thanks to the training, the agent learns to make decisions that allow it to find the object as quickly as possible without colliding. |
|