BoxStacker: Deep Reinforcement Learning for 3D Bin Packing Problem in Virtual Environment of Logistics Systems.

Manufacturing systems need to be resilient and self-organizing to adapt to unexpected disruptions, such as product changes or rapid order, in supply chain changes while increasing the automation level of robotized logistics processes to cope with the lack of human experts. Deep Reinforcement Learning is a potential solution to solve more complex problems by introducing artificial neural networks in Reinforcement Learning. In this paper, a game engine was used for Deep Reinforcement Learning training, which allows visualization of view learning and result processes more intuitively than other tools, as well as a physical engine for a more realistic problem-solving environment. The present research demonstrates that a Deep Reinforcement Learning model can effectively address the real-time sequential 3D bin packing problem by utilizing a game engine to visualize the environment. The results indicate that this approach holds promise for tackling complex logistical challenges in dynamic settings.

A One-Stage Ensemble Framework Based on Convolutional Autoencoder for Remaining Useful Life Estimation.

As the legislative pressure to reduce energy consumption is increasing, data analysis of power consumption is critical in the production planning of manufacturing facilities. In legacy studies, a machine conducting a single continuous operation has been mainly observed for power estimation. However, the production machine of a modularized line, which conducts complex discrete operations, is more like the actual factory system than an identical simple machine. During the information collection of this kind of production line, it is important to interpret mixed signals from multiple machines to ensure that there is no reduction in the information quality due to noise and signal fusion and discrete events. A data pipeline-from data collection (from different sources) to preprocessing, data conversion, synchronization, and deep learning classification-to estimate the total power use of the future process plan, is proposed herein. The pipeline also establishes an auto-labeled data set of individual operations that contributes to building an power estimation model without manual data preprocessing. The proposed system is applied to a modular factory, connected with machine controllers, using standardized protocols individually and linked to a centralized power monitoring system. Specifically, a robot arm cell was investigated to evaluate the pipeline, with the result of the power profile being synchronized with the robot program.

GadgetArm-Automatic Grasp Generation and Manipulation of 4-DOF Robot Arm for Arbitrary Objects Through Reinforcement Learning.

Automatic robot gripper system which involves the automated object recognition of work-in-process in production line is the key technology of the upcoming manufacturing facility achieving Industry 4.0. Automatic robot gripper enables the manufacturing system to be autonomous, self-recognized, and adaptable by using artificial intelligence of robot programming dealing with arbitrary shapes of work-in-processes. This paper specifically explores the chain of key technologies, such as 3D object recognition with CAD and point cloud data, reinforcement learning of robot arm, and customized 3D printed gripper, in order to enhance the intelligence of the robot controller system. And it also proposes the integration with 3D point cloud based object recognition and game-engine based reinforcement learning. The result of the prototype of the intelligent robot gripping system developed by the proposed method with a 4 degree-of-freedom robot arm is explained in this paper.

Prediction of knee adduction moment using innovative instrumented insole and deep learning neural networks in healthy female individuals.

BACKGROUND: The knee adduction moment, a biomechanical risk factor of knee osteoarthritis, is typically measured in a gait laboratory with expensive equipment and inverse dynamics modeling software. We aimed to develop a framework for a portable knee adduction moment estimation for healthy female individuals using deep learning neural networks and custom instrumented insole and evaluated its accuracy compared to the standard inverse dynamics approach. METHODS: Feed-forward, convolutional, and recurrent neural networks were applied to the data extracted from five piezo-resistive force sensors attached to the insole of a shoe. RESULTS: All models predicted knee adduction moment variables during walking with high correlation coefficients, r > 0.72, and low root mean squared errors (RMSE), ranging from 0.5% to 1.2%. The convolutional neural network is the most accurate predictor of average knee adduction moment (r = 0.96; RMSE = 0.5%) followed by the recurrent and feed-forward neural networks. CONCLUSION: These findings and the methods presented in the current study are expected to facilitate a cost-effective clinical analysis of knee adduction moment for healthy female individuals and to facilitate future research on prediction of other biomechanical risk factors using similar methods.