Smart Mobile Manipulation for Flexible Manufacturing: GoFaGO

Smart Mobile Manipulation for Flexible Manufacturing: GoFaGO Nikolas Helling1†, Francesca Negrello1†, and Manuel G. Catalano1,2† Abstract—Automation in manufacturing has rapidly expanded through the widespread integration of industrial and mobile robots. Yet many asynchronous tasks, such as machine tending, remain manual due to the technological and economic challenges of automating flexible operations with traditional robots. These limitations, combined with labor shortages, motivate the need for flexible, easy-to-program robotic solutions to automate common manual activities in manufacturing plants. This work presents a mobile manipulator platform for flexible operation integrated with a Learning-from-Demonstration (LfD) framework based on 3D perception, enabling intuitive teaching and adaptive task execution. Lab-based experiments and preliminary field tests demonstrate the feasibility of this approach for automating relevant operator-dependent tasks in various industrial settings. Index Terms—mobile manipulation, learning from demonstration, intelligent and flexible manufacturing I. INTRODUCTION AND CONTEXT The manufacturing sector has seen a widespread adoption of industrial robots on production lines and, more recently, mobile platforms for simple intra-logistics. Traditional automation, based on fixed robotic arms and conveyor-based solutions, excels in repetitive, high-volume tasks, but often struggles with flexibility. This limitation is particularly evident in common manufacturing activities such as machine tending, which are inherently asynchronous and variable. As a result, these operations remain heavily dependent on manual labor, which is costly to replace and increasingly unsustainable due to workforce shortages [1]. These challenges highlight the urgent need for innovative robotic solutions that can assist or partially substitute human workers in flexible manufacturing tasks [1]. To bridge this gap, mobile manipulators have emerged as promising candidates. Combining mobility, flexibility, and adaptability, they can operate alongside workers in dynamic factory environments [2]. However, their effectiveness is often limited by the lack of intuitive and adaptive programming methods [3], which reduces their inherent flexibility and makes them costly and impractical to deploy in real industrial settings. Current robotic machine tending applications remain confined to stationary setups and rely on complex programming or large datasets, limiting their adoption in flexible contexts [4]. While recent work on cognitive systems has shown that skill-based control can greatly enhance flexibility, such approaches still require expert-defined skills and long †All the authors are part of the JOiiNT LAB 1Soft Robotics for Human Cooperation and Rehabilitation, Istituto Italiano di Tecnologia, Via Morego, 30, 16163 Genova, Italia 2Dipartimento di Ingegneria Gestionale, dell’Informazione e della Produzione, Universit` a di Bergamo, Bergamo, Italia Fig. 1: GoFaGO’s hardware and software overview. setups [5]. This work addresses these challenges by presenting GoFaGO, a versatile mobile manipulator integrated with a Learning-from-Demonstration (LfD) framework leveraging 3D perception. Our system enables intuitive teaching of common manual tasks and adaptive execution across varying conditions, reducing the need for expert programming. We demonstrate a proof-of-concept in a lab-based industrial setting in collaboration with Cosberg, focusing on tasks identified in a feasibility study. In addition, preliminary field tests conducted with ABB validate the automation of a machine tending task across feeding stations. This work was developed at JOiiNT LAB, a joint establishment between the Italian Institute of Technology and Consorzio Intellimech. II. CASE STUDY AND PROPOSED SYSTEM A. Industrial Case Study Motivated by feedback from JOiiNT LAB’s industrial partners, we investigated robotic solutions for automating intralogistic tasks in manufacturing plants that are still manual. For this study, we considered two different contexts: Cosberg’s factories, which feature automated assembly lines for a range of components, and an ABB plant organized into robotic cells for electrical component assembly. Despite high levels of automation and the specific production layout, many repetitive operations remain manual because they occur on demand and are difficult to automate with traditional robots. Other tasks remain manual as they require the combined mobility and flexibility of human operators. We identified two representative manual tasks: Human-Machine Interface (HMI) interaction and feeding. Both are industrially relevant, as they represent low-value activities that occur in any manufacturing company and require dedicated human resources. In our preliminary study, we addressed HMI interaction in the form of restarting 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629810 181 Fig. 2: Photo-sequences of typical HMI interaction tasks (left). Overview of the lab experimental setup (right). a cycle or resetting a Cosberg machine, and cell feeding (refilling tray fixtures) within an ABB production plant. B. System Architecture 1) Hardware: The proposed system, shown in Fig. 1, is a custom mobile manipulator composed of an omnidirectional mobile base (Robotnik RB-Kairos), a collaborative robotic arm (ABB GoFa 5kg), a 3D camera (Zivid), interchangeable grippers (OnRobot 2FG7, SoftHand Industry) and onboard computing and network modules. The omnidirectional base enables navigation in constrained factory environments, while the collaborative arm ensures safe interaction with workers and equipment. The compliant gripper further increases flexibility, enabling tasks like door opening and handle turning. The platform was designed to comply with industrial safety standards, enabling future deployment in production plants. 2) Software: State-of-the-art position controllers for both the arm and base were integrated into a ROS framework, with SLAM-based navigation for factory logistics. Perception relies on 3D point cloud matching, which, together with a Learning-from-Demonstration (LfD) algorithm based on Dynamic Movement Primitives (DMPs) [6], allows to generalize trajectories from a few human demonstrations. The system interprets high-level commands (e.g., Start Machine), autonomously navigates to the target location, and identifies relevant objects and interaction points. The LfD framework then generates trajectories for each learned task, adapting to various approach configurations, essential for flexible autonomous operation. III. PRELIMINARY RESULTS The first use-case involved HMI interaction with a Cosberg assembly machine within the lab (Fig. 2). The operator first teaches the robot sequences of button and switch activations, through manual guidance. The robot could then approach the machine from a new configuration and execute the learned sequences without retraining. This is particularly relevant in manufacturing plants where HMI panels are similar across machines, but interactions remain asynchronous. Preliminary lab results confirm that the robot can quickly learn and execute common tasks such as machine start, cycle start/stop, emergency stop. The second use case regards precise machine tending and was demonstrated at an ABB production plant. The task, currently manual, involves filling tray fixtures with identical randomly placed parts taken out of a bin and feeding the tray into an assembly cell using a drawer. Operators need to interact with multiple drawers asynchronously moving across them. Using the proposed system, preliminary field tests show that bin picking can be effectively taught with few-shot demonstrations, while the mobility of the platform enables efficient navigation and feeding around the cell. IV. CONCLUSIONS AND FUTURE WORK This work demonstrates the feasibility of using a mobile manipulator combined with a DMP-based LfD framework and 3D point cloud matching to automate flexible manual tasks in industrial manufacturing. Lab experiments validated the approach for HMI interaction, enabling the robot to be easily taught how to execute asynchronous sequences without retraining. Preliminary field tests on machine tending further showed that the system can generalize few-shot demonstrations while leveraging its mobility and flexibility to handle multiple stations. Future work will focus on evaluating system robustness and reliability while continuing extended trials in production plants to evaluate real-world performance. ACKNOWLEDGMENT Special thanks to Vinicio Tincani, Alessandro Marchetti and Alessandra Paone for their technical support. REFERENCES [1] D. Romero et al. The 2024 World Manufacturing Report: New Perspectives for the Future of Manufacturing: Outlook 2030. World Manufacturing Foundation, October 2024. [2] N. Ghodsian et al. Mobile manipulators in industry 4.0: A review of developments for industrial applications. Sensors, 23(19):8026, 2023. [3] Z. Zhou et al. Advanced robot programming: A review. Current Robotics Reports, 1(4):251–258, 2020. [4] F. Jia et al. Review of current vision-based robotic machine-tending applications. The International Journal of Advanced Manufacturing Technology, 131(3):1039–1057, 2024. [5] J. Abicht et al. New automation solution for brownfield production– cognitive robots for the emulation of operator capabilities. CIRP Journal of Manufacturing Science and Technology, 50:104–112, 2024. [6] G. Lentini et al. Robot programming without coding. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 7576–7582, 2020. 182