A strategy transfer approach for intelligent human-robot collaborative assembly. (June 2022)

• Record Type: Journal Article
• Authors: Lv, Qibing; Zhang, Rong; Liu, Tianyuan; Zheng, Pai; Jiang, Yanan; Li, Jie; Bao, Jinsong; Xiao, Lei
• Is Part Of: Computers & industrial engineering. Volume 168(2022)
• Publication Date: 2022-06
• Subjects: Human-robot collaboration -- Assembly strategy -- Similarity evaluation -- Reinforcement learning -- Transfer learning; Engineering -- Data processing -- Periodicals; Industrial engineering -- Periodicals; 620.00285
• DOI: 10.1016/j.cie.2022.108047
• Languages: English
• ISSNs: 0360-8352

Abstract

Highlights: We propose the HRCA framework to improve the responsiveness of assembly system. We build the reinforcement module to realize the generation of HRCA strategy. We establish the similarity evaluation approach to measure the assembly subtasks. We design domain adaption network to achieve the cross-domain feature recognition. Abstract: In small batch and customized production, human-robot collaborative assembly (HRCA) is an important method to deal with the production demand of new-energy vehicles, which have the characteristics of rapid change and growth of personal needs. However, due to the difficulty of reusing historical assembly knowledge, it can not be used to effectively guide new tasks. Aiming at the transfer problem of collaborative strategy, this paper first defines the robot participating in the cooperation as an agent with reinforcement learning (RL) and proposes a framework of HRCA based on transfer learning (TL-HRCA). It consists of three modules: HRCA strategy generation, similarity evaluation, and strategy transfer for realizing rapid design and verification of product assembly strategy. The strategy generation module aims to establish an intelligent mapping from task to collaboration strategy based on part features. Based on the evaluation of task similarity, the mobility evaluation model divides subtasks into similar and dissimilar categories. For similar subtasks, the adversarial discriminative domain adaption is constructed to quickly design the HRCA strategy in the target domain. However, for dissimilar subtasks, the RL agent is trained continuously to obtain a new HRCA strategy. Finally, an assembly case study of power lithium batteries is conducted, of which the results have shown that TL-HRCA can improve the assembly efficiency by 25.846% compared to the traditional pre-programming assembly.