Computers & Industrial Engineering 160 (2021) 107603 Available online 5 August 2021 0360-8352/© 2021 Elsevier Ltd. All rights reserved. Deep learning-based optimization for motion planning of dual-arm assembly robots Kuo-Ching Ying , Pourya Pourhejazy , Chen-Yang Cheng * , Zong-Ying Cai Department of Industrial Engineering and Management, National Taipei University of Technology, Taipei 10608, Taiwan A R T I C L E INFO Keywords: Motion planning Optimization Dual-arm robots Random tree Long short-term memory Intelligent manufacturing ABSTRACT With the rapid technological and economic development, a growing number of companies are employing robots for their production and service operations. Motion planning is a fundamental topic in robotics that has received wide attention due to its importance in the development of industry 4.0 and intelligent manufacturing systems. This study sought to develop a deep learning-based optimization algorithm for planning collision-free trajectories of dual-arm assembly robots in complex operational environments. Given the high dimensionality of the robotic motion patterns, a Bi-directional Rapidly-exploring Random Tree integrated with the Long Short-term Memory (LSTM-BiRRT) method is proposed to enhance the effectiveness and effciency of the planning process. Numerical experiments demonstrated that the LSTM-BiRRT algorithm outperforms the state-of-the-art approaches devel- oped for motion planning of dual-arm robots in both two- and three-dimensional environments. The developed algorithm reduces the path length of the robotic operations at a signifcantly shorter computational time. The LSTM-BiRRT algorithm can serve as a strong benchmark for future developments as well as applications in the process autonomy across intelligent supply chains. 1. Introduction The recent advances in the adaptation of Industry 4.0 is facilitating the use of robots. Robots can effectively execute tasks that are dangerous, infeasible, complex, or ineffcient for the human operator (Whitney et al., 2020), and bring about operational fexibility (Richardson et al., 2018). From the wide variety of robots designed for case-specifc application areas (see (Hockstein et al., 2007)), dual-arm robots are capable of executing simultaneous processes that make them suitable for executing certain tasks (Chen et al., 2015), from complex welding operations (Gao et al., 2020), production assembly (Geismar et al., 2012) to the handling of radioactive materials in nuclear plants, and space-investigation projects (2013). Production researchers have explored the automated execution of non-coordinated and coor- dinated tasks using dual-arm robots in various contexts (Krüger et al., 2011). Some of the seminal developments in the applications of dual- arm robots are: cell production-based design of dual-arm robots in the packaging and assembly of mobile phones (Do et al., 2012); novel ap- proaches for minimizing elastic deformation and avoiding arms collision in assembling ring-shaped elastic components on cylindrical parts (Ramirez-Alpizar et al., 2014); and the motion planning algorithm integrated with RGB-D cameras, closed kinematic chain and force sen- sors for assembling furniture (Su´ arez-Ruiz et al., 2018). The design of dual- and multi-arm robots is relatively more complex than that of a single-arm robot (Basile et al., 2012; Wan & Harada, 2016) and requires motion planning models that can effectively coordinate the robotic arms operation. Robot motion planning consists of fnding a feasible path connecting the start and goal states for completing certain tasks in the workplace (Latombe, 2012). In this procedure, a controller compiles the high-level commands received by the motion planner into a set of low-level instructions required by the robot to complete a task (Farber, 2017; Ichter et al., 2018). Considering the physical limitations of the robots and the environmental obstacles they may face, the motion planner should determine the best trajectory, i.e., the shortest collision- free path towards the goal state, from the feasible alternatives (Bao & Liou, 1990). The A* algorithm developed by Hart et al. (1968) is one of the seminal works in robot motion planning that has inspired many re- searchers (Guruji et al., 2016; Lozano-Perez, 1987). Despite their merits, the early works were unable to address real-world complex situations, like physical limitations and environmental obstacles. Randomized heuristics have appeared to be effective alternatives to facilitate industry * Corresponding author. E-mail addresses: kcying@ntut.edu.tw (K.-C. Ying), pourya@ntut.edu.tw (P. Pourhejazy), cycheng@ntut.edu.tw (C.-Y. Cheng). Contents lists available at ScienceDirect Computers & Industrial Engineering journal homepage: www.elsevier.com/locate/caie https://doi.org/10.1016/j.cie.2021.107603 Received 23 January 2021; Received in revised form 1 June 2021; Accepted 2 August 2021