How AI and Robotics Can Build Furniture: A Case Study from the 2021 AI-Robot Assembly Challenge

The ”Furniture Assembly AI-Robot Challenge 2021” is a competition that utilizes artificial intelligence (AI) and robots to assemble furniture and assess the quality of the assembly. To generate commands that a robot can execute for the assembly instructions, it is crucial to develop an AI- based algorithm to recognize and interpret the assembly process based on the provided instructions. The assembly robot must be dexterous and capable of safely executing assembly tasks without operator intervention. Before the assembly process, our team employed the Faster-region-based convolutional neural networks (Faster-RCNN) and the multi-object rectified attention network (MORAN) recognition methods to identify the assembly instructions, creating a connection relationship tree structure to interpret the recognized information. The robot utilized a developed multi-fingered gripper and a manipulation station to quickly and precisely complete the assembly task. Based on these exceptional results, our team was awarded first place, thus validating the adequacy of the proposed AI-robot system for complex furniture assembly tasks. I. BACKGROUND T he field of robotics has been evolving owing to a rising demand for automated processes, but fully automated assembly remains a work in progress [1], [2]. Robots were initially introduced to unload aluminum casting machines on a Ford automobile production line. Gradually, robots have replaced humans in simple, repetitive tasks (i.e., pick and place, weld, paint) in dangerous environments due to increased degrees of freedom and payload. However, robots employed in assembly processes still require dexterous human assistance [3]. This is because realizing a fully automated assembly process with robots is challenging since parts must be placed in a specific location, making it necessary to have a designated environment. As a partial solution to the challenge of atypical assem- bly environments, Nanyang Technological University (NTU) conducted a study on the automation of the assembly of IKEA’s Stefan chair [4], [5]. The NTU team used two six-axis industrial robots and parallel grippers to connect and assemble wooden pins. They used bidirectional rapidly- exploring random tree technology to generate a collision-free assembly path and assembled the wooden pins using hybrid position/force control technology [5]. However, the NTU team only assembled a part of the chair using wooden pins, and the entire operation plan was hard-coded through engineering effort. The ”Furniture Assembly AI-Robot Challenge 2021” com- petition was held in Seoul, South Korea, to evaluate the Date of current version: 28 February 2023 performance of fully autonomous assembly robot systems. Two crucial assembly automation-related functions were eval- uated. The first was the performance evaluation of the artifi- cial intelligence (AI)-based algorithms that can automatically recognize and interpret assembly instructions to generate an assembly sequence without human intervention. The second was evaluating how well a robot system executed the entire assembly process without human assistance under atypical assembly conditions where many connector parts (pins, screw bolts, brackets) were cluttered. Participating teams were assigned two missions to evaluate their assembly automation skills in the competition. The first mission was to assemble IKEA’s Stefan chair from start to finish by following the original assembly instructions. This mission focused on judging whether the entire assembly process could be completed by manipulating the robot system without any human intervention. The second mission was to use the parts of IKEA’s Stefan chair but assemble the Stefan chair in a deformed shape rather than the original shape. The parts were identical, but the assembly instructions were modified to create a deformed Stefan chair. The participants had to use an AI-based algorithm to understand the modified instructions and generate the correct assembly sequence for the robot system to execute. Our team (Team SK2Y ) won first place in this competition due to our effective use of an AI-based algorithm and a novel robot system for the assembly process. The faster-region- based convolutional neural networks (Faster-RCNN) and the multi-object rectified attention network (MORAN) algorithms were applied to recognize the assembly instructions, and a connection relationship tree algorithm was developed to interpret the recognized commands. This AI-based algorithm quickly generated commands that the robot system could execute. Additionally, the implementation of multi-fingered grippers and a manipulation station provided additional ad- vantages during the assembly process, improving assembly time and convenience. Owing to the use of these two unique technologies, the assembly process was successfully executed with high precision and safety, and our team outperformed the others with exceptional execution time. II. FURNITURE ASSEMBLY AI–ROBOT CHALLENGE 2021 A. Competition Schedule and Rules The “Furniture Assembly AI-Robot Challenge 2021” eval- uated the combine AI’s recognition and interpretation ability with robot automation assembly technology. Four teams par- ticipated in the competition, and each team developed its own IEEE Robotics & Automation Magazine (RAM) paper, presented at ICRA 2023, London, UK. Cite as RAM paper. IEEE Robotics & Automation Magazine (RAM) paper, presented at ICRA 2023, London, UK. Cite as RAM paper.