Conceptualisation and Implementation of a ROS-Based Robotic Cell for a Flexible Pre-Assembly Task

Abstract

Modern manufacturing is currently shifting toward highly flexible, high-mix, and low-volume production cycles, requiring small and medium-sized enterprises (SMEs) to adopt reconfigurable automation to remain competitive. However, the adoption of such technologies is often slowed down by the high cost and rigidity of commercial solutions, which typically rely on proprietary toolchains and necessitate specialized expert knowledge for reconfiguration. This study proposes a methodological framework for a modular robotic cell based on an open-architecture approach using ROS2 middleware, designed to be maintained by personnel without deep robotics expertise. The methodology emphasizes the replacement of fixed mechanical fixtures with an AI-driven perception pipeline, utilizing YOLO-based image segmentation to enable the autonomous localization of heterogeneous components. A rigorous tolerance chain analysis defines the design requirements of custom 3D-printed self-aligning fingertips, providing a mathematical and mechanical basis for ensuring assembly feasibility under tight geometric constraints. By adopting a node-based software topology, the framework facilitates rapid task reconfiguration and hardware interoperability. Experimental validation in an industrial-like environment confirms that this integrated approach provides a scalable pathway with the potential to improve cost-effectiveness in high-mix low-volume production scenarios to overcome manual production bottlenecks through intelligent, reconfigurable automation.

1. Introduction

Modern manufacturing is currently transitioning from traditional mass production models toward high-mix, and low-volume production cycles [1]. This transformation is driven by rapidly changing market demands and an increasing requirement for customised product variants. Within this context, small and medium-sized enterprises (SMEs) face the critical challenge of maintaining economic viability while reacting effectively to frequent product launches. While industrial robotics has historically succeeded in repetitive, high-volume tasks, conventional commercial solutions often rely on monolithic programming and are designed for product-specific applications, making the time and cost associated with reconfiguration too prohibitive for small-lot-size production.

Although advanced commercial automation solutions already exist in the market, they are frequently tightly coupled to proprietary toolchains that remain expensive and platform-dependent. A major obstacle for SMEs approaching process automation is that setting up and reconfiguring these existing solutions usually requires expert knowledge and significant time for testing and fine-tuning. Because SMEs often lack this specialized expertise in-house, they avoid introducing such solutions even when they are economically justifiable. The ability to reconfigure robotic cells flexibly and independently using open systems by personnel without specialized expertise remains a significant technical and operational challenge.

To address these limitations, the emerging Industry 4.0 and Industry 5.0 agendas emphasize the development of reconfigurable robotic workcells that diverge from fixed setups by adopting an open architecture approach [1]. This design focus prioritizes modularity in both hardware and software to facilitate rapid adaptation to new tasks and workflows [2,3]. Central to this flexibility is the replacement of rigid mechanical fixtures with sensor-driven systems, e.g., by utilizing advanced computer vision and force-torque feedback, allowing robots to handle a wider variety of objects with different geometric properties.

The Robot Operating System (ROS), and its evolution ROS2 (https://docs.ros.org/en/humble/index.html accessed on 1 July 2026), has emerged as the leading middleware solution to bridge the gap between academic robotics research and industrial application [4]. By providing a hardware-agnostic abstraction layer, ROS enables the creation of a modular software architecture where individual functional entities, such as vision processing, motion planning, and peripheral control, operate as independent nodes. This node-based structure allows for dynamic graph modifications and can further simplify the debugging process. The popularity of ROS is further supported by a growing ecosystem of industrial drivers and open-source projects like MoveIt, which facilitate complex tasks such as collision-aware path planning and inverse kinematics [5].

This article addresses the challenge of implementing a flexible robotic pre-assembly cell where random bulk delivery and minimal geometric variances of components to be assembled typically necessitate specialized expert knowledge for system reconfiguration. An overview of the state of the art and the research objectives is provided in Section 2. Then, Section 3 goes on to describe the identified modular robotic work-cell utilizing an open-architecture approach designed to be managed by personnel without deep robotics expertise. Experimental validation was conducted in the Smart Mini Factory of the Free University of Bozen-Bolzano, providing conditions representative of real industrial manufacturing scenarios, and it is presented in Section 4. Finally, Section 5 draws conclusions and outlines potential future developments.

2. Background and Related Works

The growing demand for customized products, shorter product life cycles, and reduced production volumes demands increasingly flexible and reconfigurable production systems which, unlike traditional automation solutions, can be optimized for low-volume and high-mix productions [6]. This challenge is particularly relevant for SMEs [7], where the development of dedicated automation solutions may not be economically sustainable due to high engineering and commissioning costs. In this context, open and modular robotic frameworks have emerged as key enablers for the implementation of adaptable industrial solutions supporting low-volume high-mix production scenarios [8].

2.1. ROS-Industrial for Manufacturing Automation

As previously introduced, ROS has become one of the most widely adopted software frameworks for robotics research and development [9], providing a distributed peer-to-peer communication infrastructure that enables the integration of heterogeneous software components through a modular architecture [10]. Unlike previous monolithic robotic software solutions, ROS promotes software reusability, scalability, and interoperability among different hardware and software platforms, facilitating the eventual development of complex robotic applications [9].

Despite its success in research environments, the adoption of ROS in industrial applications initially faced several challenges related to robustness, standardization, long-term maintenance, and compatibility with industrial hardware [11]. To address these limitations, ROS-Industrial (ROS-I) was introduced with the objective of bridging the gap between robotics research and industrial automation [12]. ROS-I indeed aims to extend the ROS ecosystem by providing standardized robot drivers, industrial communication interfaces, sensor integration modules, and software components specifically thought for ROS-based integration into industrial environments. The ultimate objective of ROS-I can be seen as the combination of the flexibility and innovation potential of academic research with the robustness and reliability requirements of industrial production systems [11] by providing interoperable robotic interfaces across heterogeneous platforms.

Recent research demonstrates the increasing adoption of ROS-based architectures in industrial automation. In [13] for instance, the authors developed a sensor-guided robotic assembly framework integrating multiple vision systems, force/torque sensing devices, and motion planning algorithms through a ROS-based architecture. Similarly, Qu et al. [14] proposed a ROS-based robotic sorting system integrating computer vision, object detection, motion planning, and robotic manipulation for real-time grasping applications. More recently, Kawshan et al. [2] demonstrated the integration of Unity and ROS within a Digital Twin framework for robot path planning and real-time execution, highlighting the role of ROS as an enabling middleware for Industry 4.0 industrial robotic systems.

These developments demonstrate that ROS-I has evolved from a research-oriented framework into a platform capable of supporting advanced industrial robotic applications requiring modularity, interoperability, and rapid system integration [15]; despite this, existing ROS-based solutions are predominantly oriented towards generic pick-and-place tasks, collaborative workstations, or large-scale assembly processes.

2.2. Flexible and Reconfigurable Robotic Assembly Systems

Reconfigurable Manufacturing Systems (RMS) were introduced to overcome the limitations of traditional production lines by enabling rapid adaptation to changes in product design, production volume, and process requirements [16,17]. In parallel, the Industry 4.0 paradigm promoted the integration of cyber-physical systems, bi-directional communication, and intelligent decision-making within industrial environments [18,19]. To address these challenges, flexible robotic assembly systems based on modular hardware and software architectures have been proposed in recent academic research. In [20], for instance, a self-adaptive robotic assembly system based on ROS and MoveIt is developed, integrating vision-based object recognition and collision-free motion planning to operate in semi-structured industrial environments, demonstrating the benefits of adaptive robotic architectures in reducing reprogramming effort and enhancing flexibility.

Gaspar et al. [21] proposed a reconfigurable robotic workcell specifically targeting SMEs and few-of-a-kind production scenarios. Their solution combined modular hardware components with a ROS-based software framework capable of supporting rapid hardware and software reconfiguration, significantly reducing setup times and engineering costs. Similarly, Ewert et al. [1] introduced a ROS-based robotic cell incorporating autonomous assembly planning capabilities, enabling the generation of assembly sequences directly from product descriptions and production constraints.

More recently, Nemec et al. [3] presented a flexible robotic workcell capable of operating autonomously or in collaboration with human operators. Their system integrated learning-based adaptation strategies, compliant control, and human intention recognition within a ROS-based architecture, highlighting the growing convergence between flexible automation and collaborative manufacturing. In addition, Terei et al. [22] demonstrated the applicability of ROS-based architectures in industrial micro-assembly scenarios, confirming the potential of open and modular solutions for highly specialized assembly processes.

2.3. Vision-Guided Robotic Assembly and Manipulation

While modular software architectures and reconfigurable robotic cells provide the foundation for industrial flexibility, effective operation in highly variable production environments also requires advanced reliable perception in detecting, localising, and manipulate components under uncertain positioning and orientation conditions [23]. Vision systems have therefore become a key enabling technology for robotic assembly, and computer vision technologies are increasingly integrated to reduce the dependence on dedicated fixtures and highly structured production environments [24]. By providing real-time information about component location, orientation, and process status, vision systems enable robots to adapt their actions to changing operating conditions and product configurations.

Several studies have investigated the integration of vision systems within robotic assembly frameworks. Herrero et al. [25] proposed a skill-based approach for robotic vision applications, demonstrating how reusable perception modules can support different assembly processes while improving software reusability. Similarly, Lin et al. [26] developed a vision-guided robotic system for the handling of small mechanical components, combining object localization and orientation estimation to support automated manipulation tasks.

Vision-guided manipulation has also been applied to more complex assembly and handling scenarios. Chen et al. [27] proposed a robotic perception framework integrating object recognition, pose estimation, and grasp planning, while Qu et al. [14] developed a ROS-based robotic system combining visual perception, motion prediction, and grasp execution for dynamic manipulation tasks, demonstrating the importance of integrating perception capabilities to improve the robustness and adaptability of robotic operations. Lastly, other applications also explore quality inspection and process verification, supporting higher levels of automation and reducing manual intervention [28].

Overall, the reviewed literature confirms the growing maturity of ROS-based industrial robotics, the increasing adoption of reconfigurable manufacturing concepts, and the important role of perception technologies in enabling adaptive robotic manipulation. Nevertheless, the integration of these capabilities into unified robotic systems capable of supporting flexible assembly operations under low-volume high-mix production conditions remains an open research challenge. In particular, there is still a need for robotic cells that combine modularity, reconfigurability, and autonomous perception while maintaining the simplicity, adaptability, and cost-effectiveness required by SMEs [29].

2.4. Research Objectives and Contribution

As mentioned so far, commercial automation solutions exist, but they often rely on proprietary rigid toolchains that require platform-specific expertise. For SMEs, the ability to reconfigure such cells flexibly by staff without deep specialized knowledge in robotics represents a significant edge to adopt process automation. To address these gaps, this study proposes a reconfigurable framework designed for high-mix, low-volume hardware assembly. The specific technical objectives are:

• Open-Architecture Implementation: leverage open-source middleware to create a modular software–hardware ecosystem that supports rapid task reconfiguration.
• AI-Driven Perception: integrate a deep learning-based image segmentation pipeline that simplifies the creation of a customised dataset, the training of the model and its deployment into production.
• Minimal hardware reconfiguration: minimise the hardware changes required to reconfigure the system by utilising widely adopted technical solutions.

These objectives have guided the development of this work and of the functional design concept and system configuration presented in the following section.

3. Functional Design Concept and System Configuration

3.1. Methodological Framework for Rapid Reconfiguration

The proposed approach addresses one of the main barriers preventing the adoption of flexible robotic assembly systems in SMEs: the complexity of reconfiguring both software and hardware when production requirements change. Rather than developing a task-specific solution, the proposed framework combines an open software architecture, AI-based perception, and modular hardware components to simplify system adaptation while limiting engineering effort. As illustrated in Figure 1, the coordination module orchestrates the interaction between perception, manipulation, gripping, and feeding modules, each encapsulating a specific functionality. This modular decomposition allows individual subsystems to be replaced or extended without affecting the overall control logic.

The framework further promotes rapid deployment through an image-segmentation pipeline based entirely on open-source tools, enabling non-specialist users to create custom datasets, train deep-learning models, and deploy them within the robotic cell using affordable computing resources. At the hardware level, the adoption of commercially available devices combined with rapidly manufacturable 3D-printed tooling minimizes the physical modifications required when introducing new products. Although expert intervention is still required for complex manipulation strategies and system optimization, the proposed methodology significantly reduces the engineering effort needed for reconfiguration. The following sections describe the implementation of this framework through a representative case study involving manual pre-assembly operations for early-stage electrical cabinet wiring.

3.2. Case Study

This section summarizes the case study that inspired this work. The assembly process for electrical switchboards and cabinets, particularly in the case of complex units, remains a largely manual task today due to the need for careful routing of cables and connectors, as shown in Figure 2.

There are, however, certain preliminary tasks that offer great potential for automation and mechanisation. Often, the cabinet assembly cycles begin with the creation of sub-assemblies and pre-assembled mechanical components that will form the framework and structure upon which the electrical circuit will be built. Although these are mechanical assemblies with a small number of components, they exhibit high variability as they are closely dependent on the final product. Moreover, they are characterised by small size components, as they don’t need to occupy too much space in the cabinets. Traditional automation solutions aim to address each sub-assembly case individually, involving development costs that are difficult for SMEs to sustain. The primary objective of this case study is therefore to develop a robotic cell capable of ensuring high flexibility and reconfigurability to cope with the low-volume high-mix nature, and small-size components. As specific case study, the proposed system is designed to produce a pre-assembled unit consisting of a central primary screw upon which various elements are sequentially stacked and inserted in a predefined order. The manipulated workpieces are categorized into five distinct groups, characterized by varying physical and geometric properties (see Figure 3):

• Hexagonal head screws;
• Contact washer;
• Slightly conical washers;
• Cylindrical head screws;
• Bronze spacers.

Figure 3. Technical illustration of the pre-assembly considered as case study.

Figure 3. Technical illustration of the pre-assembly considered as case study.

Standard industrial supply conditions dictate that components number 1 through number 4 are delivered in bulk within plastic or cardboard containers, resulting in completely random orientations. While bronze spacers (number 5) are typically supplied resting on their longitudinal axis.

3.3. Hardware Configuration

As shown in Figure 4, the pivot point of the entire cell is a collaborative robotic platform: the Doosan M0609 (Doosan Robotics Inc., Seongnam-si, Republic of Korea, https://www.doosanrobotics.com/en/product-solutions/product/m-series/m0609/ accessed on 1 July 2026). It is a 6-axis collaborative robotic platform (cobot) designed for high-speed, high-precision tasks in small workspaces. It features a 6 kg payload capacity and a 900 mm reach. Maximum joints’ speeds are reported in

Table 1. Doosan M0609 joints’ parameters.

Joint		Joint Range	Max Joint Speed
Joint 1	Base	$\pm {360}^{\circ }$	${180}^{\circ }$/s
Joint 2	Shoulder	$\pm {360}^{\circ }$	${180}^{\circ }$/s
Joint 3	Elbow	$\pm {150}^{\circ }$	${180}^{\circ }$/s
Joint 4	Wrist 1	$\pm {360}^{\circ }$	${225}^{\circ }$/s
Joint 5	Wrist 2	$\pm {360}^{\circ }$	${225}^{\circ }$/s
Joint 6	Wrist 3	$\pm {360}^{\circ }$	${225}^{\circ }$/s

.

There are two main reasons that support the choice of this machine. First of all, this robotic platform is equipped with 6 high-performance torque sensors. They allow for hybrid force-position control integration which comes convenient when dealing with assembly operations, where small interferences between components may be required. Additionally, the high-tech torque sensors allow 5 out of 6 axes to rotate a full ${360}^{\circ }$, and let the robot reach a repeatability of ±$0.03$ mm. Second of all, this robotic platform is ROS2 supported by the manufacturer. The choice of an off-the shelf ROS2 compatible robot is dictated by the principle of open architecture suggested by ROS-I that characterises the aim of this work, facilitating a modular and reconfigurable environment that contrasts with traditional, rigid assembly systems.

Two additional devices were integrated with the Doosan to provide the ability of sensing the surrounding environment and to interact with it. Firstly, a ROS2 compatible tooling was attached to end of the arm. Specifically, the choice was the OnRobot 2FG7 (OnRobot A/S, Odense, Denmark, https://onrobot.com/en/products/2fg7-finger-gripper accessed on 1 July 2026): a compact, high-performance all-electric parallel gripper designed for tight production environments. This unit eliminates the need for compressed air while delivering precise and programmable control over gripper kinematics and dynamics. The system offers an adjustable gripping force ranging from 20 N to 140 N with a ±5 N tolerance, a total mechanical stroke of 38 mm, and an operational payload capacity of up to 11 kg for form-fit grasping topologies or 7 kg for force-fit applications. Sensor integration within the onboard controller supports closed-loop feedback, enabling real-time grip verification and lost-grip detection with a precision repeatability of ±$0.1$ mm. Even though this gripper has no OnRobot official ROS2 drivers, community developed drivers to bridge this gap ensuring a flawless integration into the cell, and creating a solid base for future ROS-I applications.

The second device is an industrial CCD camera: a Basler ace acA2040-25gc (Basler AG, Ahrensburg, Germany, https://www.baslerweb.com/en/shop/aca2040-25gm/ accessed on 1 July 2026) coupled with a Kowa LM6JC (Kowa Optimed Deutschland GmbH, Düsseldorf, Germany, https://www.kowavision.com/products/lm6jc?srsltid=AfmBOophUdv9DDRq0KCQe-esNttadpJMjIqUF7dHAPivcYw5X0-xldzK accessed on 1 July 2026) high-resolution machine vision lens. This camera is built around a CMV4000 CMOS 1 inch sensor, that yields to a $2048\times 2048$ resolution and ($4.2$ Megapixels). To facilitate high-speed acquisition, the sensor employs a progressive-scan global shutter mechanism with a pixel size of $5.5\times 5.5\mathsf{\mu }\mathrm{m}$, achieving a native throughput of 25 fps at full resolution. Data transmission and power delivery are managed concurrently via Gigabit Ethernet utilizing the GigE Vision standard and Power over Ethernet (PoE IEEE 802.3af), housed within a compact $42 \times 29 \times$ 29 mm enclosure weighing 90 g. This camera ensures high performances in a small footprint. Optically matched via a standard C-mount interface, the Kowa LM6JC lens provides a fixed wide-angle focal length of 6 mm. The lens features completely manual iris and focus rings, operating across an adjustable aperture range of $f/1.4$ to $f/16$, with a minimum object distance of 100 mm. Pylon camera driver for ROS2 are officially provided as an open-source project by Basler, creating de facto a ROS-I ready device. Moreover, this camera-lens configuration provides a short-exposure vision pipeline suitable for integration into the robotic cell.

The last piece of equipment included in the cell is a feeding system. We implemented the ARS Automation FlexiBowl 500 C (ARS Automation S.r.l., Civitella in Val di Chiana, Arezzo, Italy, https://www.flexibowl.it/flexibowl#fb-500 accessed on 1 July 2026), a rotating parts-feeding platform designed to handle heterogeneous components with complex geometries and varying material compositions. Operating on a combined mechanism of bidirectional rotary actuation and impulsive flip-vibration, the system prevents part entanglement and optimizes distribution across a total surface load capacity of up to 7 kg. The device is geometrically optimized for components measuring between 5 mm and 50 mm with individual masses under 100 g. Physically, the feeder features an overall diameter of ⌀640 mm, a designated picking height of 270 mm, and a total system mass of 42 kg. To maximize compatibility with computer vision systems, it incorporates an integrated backlit illumination window. It requires both electricity (for the rotation) and pneumatic (for the flip) supplies. Because the manufacturer does not supply an official ROS2 support, a custom, open-source ROS2 hardware driver was developed in-house to integrate the feeder into the cell. Following in the development the ROS-Industrial principles of modularity, and standardized communication, the drivers establish a socket communication interface over standard TCP/IP-UDP protocols, abstracting the device’s native parametrized routines (such as flip, shake, and precise angular disc indexes) into standard ROS2 services and actions. We decided to mount these devices on an Isel USA Inc. (Hicksville, NY, USA) industrial workbench. This bench features all aluminium construction (except for the hardware and mounting feet) and has a T-slot tabletop. The T-slots allow for quick and easy fixturing and adjusting, ensuring an easy reconfiguration of the robotic cell. As shown in the system layout, the M0609 is placed at the back of the table to ensure adequate clearance for the safe loading and unloading of the components. A detailed representation of the cell network is shown in Figure 5.

Each device has its own controller, which is connected via Ethernet to a D-Link DGS-1100-08P Gigabit Network switch (D-Link (Europe) Ltd., London, United Kingdom). The only exception is the camera, which is directly connected to an external workstation, where also the ROS2 controller runs. We chose a workstation equipped with an Intel Xeon W-2125 CPU, 32 GB of high-speed DDR4 RAM and an NVIDIA GeForce RTX 2080 Ti graphics processing unit (GPU). Built on the Turing (TU102) microarchitecture, this graphics processor features 4352 CUDA cores dedicated to matrix mathematics, and 11 GB of GDDR6 VRAM. This high-performance computational hardware tier acts as the central hub for the custom ROS2 node topology, as well as for the deep-learning-based part localization pipeline. Once finalized, the mechanical and network systems used for our experiments were prototyped, as shown later in the experimental setup (see Section 4).

3.3.1. Chain of Tolerances

In the design of automated assembly cells, the definition of a rigorous tolerance chain is fundamental to ensure process reliability. The primary objective of this analysis is to determine the feasibility of the assembly task, by evaluating the cumulative effect of geometric variance and mechanical repeatability. This analysis is particularly critical in the context of the suggested case study, where the assembly task requires the precise insertion of washers and spacers through the body of the screw. To determine whether the proposed robotic system can successfully execute these tasks without mechanical interference, a mathematical model can be developed. The analysis begins with the definition of most critical dimension for each part, i.e., the internal diameter, if it’s a receiving element, or the external diameter, if it’s an insertion element. We therefore define:

• Internal diameter of the spacer $d_{i,e}$;
• Internal diameter of the washers $d_{i,w}$;
• External diameter of the screw thread $d_{e,s}$.

In a robotic assembly context, the available radial clearance must accommodate the relative radial displacement between the centres of the mating components to avoid mechanical jamming during placement. The relevant clearance $c_r$ is therefore defined as half the difference between the internal diameter of the receiving component and the external diameter of the inserted component:

$$c_r={\displaystyle \frac{min(d_{i,e},d_{i,w})-d_{e,s}}{2}}$$

(1)

Under a conservative worst-case assumption, assembly feasibility requires:

$$c_r>r_{p,s}+r_{p,w}$$

(2)

The positioning error of the screw $r_{p,s}$ is conservatively expressed as the sum of the robot’s movement repeatability $r_r$, which refers to the precision with which a robot can return to a specific programmed coordinate in the workspace, and the mechanical variance in how the screw is seated within the gripper $r_{g,s}$.

$$r_{p,s}=r_r+r_{g,s}$$

(3)

Similarly, the positioning error of the washer $r_{p,w}$ is a function of $r_r$ and $r_{g,w}$

$$r_{p,w}=r_r+r_{g,w}$$

(4)

By substituting Equations (3) and (4) into Equation (2), we derive the second feasibility criterion, which allows to solve for the maximum allowable positioning error of the receiving part, i.e., $r_{p,w}$.

$$r_{p,w}<c_r-2r_r-r_{p,s}$$

(5)

It is important to note that $r_r$ appears twice in Equation (5) because the robot’s precision affects both the pick-up pose and the final placement pose. Figure 6 shows graphically the two feasibility criteria.

Using the nominal dimensions reported in

Table 2. Nominal dimensions used in the chain of tolerances computation.

Internal diameter of the spacer	$d_{i,e}$	13 mm
Internal diameter of the washers	$d_{i,w}$	13 mm
External diameter of the screw thread	$d_{e,s}$	12 mm
Robot repeatability	$r_r$	$0.1$ mm
Gripping repeatability of the screw	$r_{g,s}$	$0.1$ mm

, the available radial clearance is $0.5$ mm. The resulting maximum allowable error for the washer $r_{p,w}$ is $0.2$ mm. This requirement guided the design of the custom self-centring fingertips described in Section 3.3.2.

3.3.2. Gripper Design Proposal

Starting from the tolerance constraints derived in Section 3.3.1, the design of the gripper and in particular of its fingertips was directly driven. The allowable positioning error for the receiving component is limited to $0.2$ mm, which imposes a very tight request on the consistency with which components are held and centred by the gripper. In this context, the fingertip design was conceived as a geometrically self-aligning system, thereby ensuring compliance with the feasibility criteria defined in Equations (2) and (5). As shown in Figure 7a, the proposed solution adopts the gripper mentioned in Section 3.3 with custom fingertips specifically designed to closely match the geometries and sizes of the objects to be gripped. The inner curvature increases the contact area and promotes repeatable positioning of screws within the gripper, contributing to lowering $r_{g,s}$. parts within the gripper, contributing to lowering $r_{g,s}$ and $r_{g,w}$. In addition to the curved gripping surfaces, a central cavity is introduced between the fingers to address the most critical component in the tolerance chain, namely the washer, whose positioning error directly affects the satisfaction of the feasibility condition. This cavity passively guides ring-shaped components toward the centre of the gripper during the closing phase, effectively implementing a mechanical self-centring mechanism. Furthermore, the external geometry of the fingertips was designed with smoothed and rounded edges, allowing the gripper to interact safely with partially cluttered environments by pushing away interfering elements. The overall structure is compact, reducing lever arms and potential compliance effects that could degrade repeatability during manipulation.

As shown in Figure 7b, the fingertips represent the only hardware element that must be adapted or replaced to effectively cope with a low-volume, high-mix production scenario, enabling rapid reconfiguration while preserving system accuracy. From a design philosophy perspective, the chosen fingertip solution reflects a preference for geometric coupling and passive compliance rather than relying on active gripping technologies such as suction or electromagnetic systems, which were considered less suitable due to limitations in precision, material compatibility (e.g., non-magnetic bronze spacers), or reliability when dealing with small parts.

To validate whether the designed fingertips effectively reduce gripping-induced variability and satisfy the tolerance requirements, an experimental campaign was carried out focusing on the repeatability and robustness of the grasp. Components with external diameters ranging from 12 mm, corresponding to the screw, to 25 mm, corresponding to the brass spacer, were used. Although the current geometry provides an available opening of approximately 35 mm, grasping reliability beyond the tested dimensional range was not assessed. As shown in Figure 8, 30 grasping trials were performed for each component type, with the parts deliberately positioned in known and controlled configurations to isolate the performance of the gripper from perception errors. Each trial included not only the closure of the gripper on the component but also the lifting and handling phases, thus evaluating the stability of the grasp throughout the manipulation cycle. The results showed that no grasp failures or part drops occurred across all tested components, indicating that the fingertip geometry provides sufficient constraint and repeatability to maintain consistent positioning. This outcome supports the assumption that the adopted design can effectively limit the additional uncertainty introduced by the gripping process, allowing the system to remain within the error bounds imposed by the tolerance chain analysis. In this sense, the gripper design can be interpreted not merely as a handling device, but as an integral component of the tolerance management strategy, where mechanical design choices are explicitly used to control and minimize uncertainty contributions, ultimately enabling the successful execution of the automated assembly task under the tight geometric constraints identified in the previous section.

3.4. SW Configuration

In this work, a complete automated assembly system was conceived and implemented. The overall system integrates several functional modules, including component recognition, manipulation, robot motion control, and system coordination modules. In the following subsections, the different system modules are explained in more detail. Specific emphasis is given to the vision module and the vision pipeline in general.

3.4.1. Modular Approach

The software architecture adopted in this work follows a modular approach, with the objective of ensuring flexibility, scalability, and ease of reconfiguration in a low-volume, high-mix production context. In contrast to monolithic control systems, where all functionalities are embedded in a single control layer, the proposed solution decomposes the overall application into independent yet interconnected modules. Coordination between the software modules is illustrated in Figure 9. Overall, the architecture follows the principles of flexibility and modularity consistent with the ROS-Industrial approach, and it is divided into five software modules:

• Coordination module;
• Robot motion module;
• Feeding module;
• Gripping module;
• Vision module.

Figure 9. System architecture of the modular robotic cell control framework.

Figure 9. System architecture of the modular robotic cell control framework.

The coordination module acts as the central orchestrator of the assembly process, managing the execution flow and coordinating interactions among the perception, feeding, manipulation, and control subsystems. Implemented as a ROS2 node, the module initializes the process parameters, including robot poses, assembly configurations, and the bill of materials (BOM), and supervises the execution of the complete assembly cycle. The coordination logic is based on a finite state machine (FSM), whose states represent the main phases of the process: component distribution, robot positioning, object detection, picking and placement, part repositioning, and task completion. The FSM ensures deterministic execution by activating the appropriate subsystem according to the current process state and the outcome of previous operations. During execution, the module acquires component detections from the vision system, selects the target part according to the BOM, and coordinates the robot and gripper actions required for grasping and placement. If no suitable component is detected, the feeding system is activated to redistribute the parts and a new detection cycle is initiated. The module also manages execution timing, cycle counting, and process monitoring until all components have been assembled. Communication with the other software modules is implemented through ROS2 services and actions, enabling modular, asynchronous, and scalable integration. The logical sequence of the FSM states is illustrated in Figure 10.

The robot motion module is built upon the ROS2 library provided by Doosan Robotics and is responsible for executing all manipulator movements required during the assembly process. The module supports both joint-space and Cartesian-space motion commands, enabling the robot to perform operations such as homing, object approach, placement, and trajectory retreat. In addition, compliant motion capabilities are available during contact-sensitive tasks, allowing safer and more reliable interactions with the assembly fixture and components. By abstracting low-level robot control, the module provides a standardized interface that enables the coordination module to execute complex manipulation sequences through high-level commands.

The feeding module provides the interface between the robotic cell and the Flexibowl flexible feeding system. Its primary role is to control part distribution and presentation, ensuring that components are continuously rearranged to facilitate reliable detection and grasping. The module exposes feeding operations such as rotation, shaking, and flipping through dedicated ROS2 actions, allowing these functions to be executed asynchronously and coordinated with the other subsystems. By abstracting the device-specific communication layer, the module enables seamless integration of the Flexibowl within the ROS2 architecture and provides a standardized mechanism for triggering feeding actions during the assembly process.

The gripping module provides the interface between the robotic cell and the adaptive gripper used for component manipulation. Based on an extended version of the OnRobot ROS2 driver, the module enables high-level grasping operations while ensuring seamless integration with the overall software architecture. It supports both internal and external gripping strategies and provides real-time feedback on gripper status and opening width. By abstracting device-specific communication and control, the module allows the coordination system to execute grasping tasks through a standardized set of services.

The vision module is responsible for detecting, classifying, and localizing the components required for the assembly process. Image acquisition is performed through an industrial Basler camera, while perception is based on a customized You Only Look Once (YOLO) segmentation model integrated within a ROS2 processing pipeline. The module extracts object attributes relevant to manipulation, including class, position, orientation, and confidence level, and makes this information available to the coordination and manipulation modules. Its modular architecture enables reliable operation under varying conditions and supports the autonomous execution of the assembly workflow. The following sections describe the vision pipeline and the training procedure of the YOLO model in greater detail.

3.4.2. Dataset Creation

The creation of the dataset adopted for training the vision module was carried out using Computer Vision Annotation Tool (CVAT), a widely used platform for image annotation that can be deployed in a self-hosted configuration at no licensing cost. This aspect was particularly advantageous, as it enabled full control over data management and annotation workflows while maintaining compatibility with modern deep learning pipelines. The dataset was composed of images acquired directly from the industrial setup, using the same camera employed in the final system, ensuring consistency between training and deployment conditions. In total, the dataset included 250 labelled samples, distributed across training, validation, and test sets, following a 70-20-10 split.

The class distribution was defined to reflect the components involved in the assembly process, specifically including multiple washer types and orientations, cylindrical spacers, and screws. Each class corresponds to a distinct geometric and functional category, and particular care was taken to ensure that all classes were adequately represented despite inherent differences in visual complexity and occurrence frequency. With reference to Figure 11, the annotation process was performed manually in CVAT using instance segmentation masks, enabling the extraction of both shape and spatial features required for downstream tasks. Also partially captured objects were labelled as full, to increase the chances of a positive recognition in case of partial cluttering. Special attention was given to classes such as washers, which exhibit low contrast and limited geometric features, making them more challenging to annotate consistently. To improve model robustness and generalization, images were collected under varying illumination conditions, including low, medium, and high lighting scenarios, as well as different component arrangements. From a workflow perspective, CVAT allowed efficient organization of annotation tasks, collaborative labelling if required, and direct inspection of annotation quality. A key advantage of the selected tool is its ability to export annotated datasets directly in formats compatible with YOLO-based frameworks, including segmentation-compatible annotations. Overall, even though well structured, the dataset creation process is the other application-specific element.

3.4.3. Image Segmentation

The image segmentation module is based on the YOLO framework, a state-of-the-art unified architecture that reframes object detection as a single regression problem. Unlike traditional multi-stage pipelines that employ sliding windows or region proposal methods, YOLO evaluates the entire image in a single pass through a convolutional neural network. The system divides the input image into a grid, where each cell is responsible for predicting bounding boxes, confidence scores, and class probabilities simultaneously. This global reasoning approach ensures high speed and significantly fewer background errors compared to architectures like Fast R-CNN. This allows the network to extract detailed shape information, which is essential for downstream tasks such as orientation estimation and grasp point computation [30]. YOLO was selected for this application primarily due to its real-time processing capabilities and its ability to learn generalizable representations of objects, which is critical for industrial environments with fluctuating lighting. While standard object detection provides bounding boxes, image segmentation is implemented to provide pixel-level masks, identifying the precise contours and centroids of components on the vibratory feeder. This pixel-level precision is fundamental to navigating the strict sub-millimetric tolerance chain required for automated insertion. The trained YOLO model is integrated into the ROS2-based vision module through a dedicated perception node, acting as an action server that returns detection results on request. When triggered by the coordination module, the node receives images from the Basler camera driver, converts them to an OpenCV-compatible format, and performs inference. The resulting segmentation masks are post-processed to extract geometric features, including centroids, bounding polygons, orientation angles, and grasp points according to the detected class. Cylindrical components are analysed using minimum-area bounding rectangles to estimate orientation, while elongated parts such as screws are processed to identify grasp points along their principal axis. The extracted information is converted into metric coordinates in the camera reference frame and returned as structured output, including class identifiers, confidence scores, and pose representations. This enables seamless integration with the manipulation pipeline, while the ROS2 action interface supports asynchronous, non-blocking execution to maintain system responsiveness.

4. Prototyping and Preliminary Experimental Tests

The suggested design has been implemented and a preliminary experimental validation was carried out at the Smart Mini Factory laboratory of the Free University of Bozen-Bolzano, where a dedicated robotic cell was set-up to replicate the targeted assembly scenario. The complete architecture described in the previous sections, i.e., comprising the vision, coordination, motion, manipulation, and feeding modules, was fully deployed and integrated within a ROS2-based framework, ensuring seamless communication between hardware and software components. The physical setup included the equipment listed in Section 3.3, while the software distributions, packages, and versions used for the implementation are reported in Appendix A. Particular attention was given to the integration of the system components, ensuring that the modular design principles could be effectively translated into a working prototype. This implementation phase was essential to validate not only the individual modules, but also their interaction within a real industrial-like environment. The following sections present the preliminary experimental tests conducted to assess system performance and robustness.

4.1. Training of the Neural Network

The image segmentation model was trained using a supervised learning approach based on the YOLO26 architecture. In supervised training, manually annotated images are used to iteratively optimize the network parameters by minimizing a multi-objective loss function that jointly considers object classification, localization accuracy, and segmentation mask quality. Due to hardware limits, the training process was performed for 120 epochs with an input image size of $512\times$ 512 px and a batch size of four images. To improve generalization and reduce overfitting, several data augmentation techniques were applied, including horizontal and vertical flipping, small rotations, scaling, and HSV colour perturbations. Label smoothing was set to $0.2$, while early stopping with a patience of 50 epochs was adopted to prevent unnecessary training once validation performance saturated. A cosine learning-rate schedule and automatic mixed precision (AMP) were also enabled to improve optimization efficiency and reduce computational costs.

At the time of this study, YOLO26 represented the most recent model generation available. Adopting the latest architecture is advantageous because it typically incorporates improvements in feature extraction, optimization strategies, and inference efficiency, resulting in better accuracy–speed trade-offs. Ultralytics implementation version 8.4.101 was used as software framework to run YOLO26-seg, which is distributed in different sizes, such as Nano (n), Small (s), and Medium (m). They are the only sizes that the graphics card used for training (Nvidia RTX 2080 Ti) supported. A comparison among them is reported in

Table 3. Size and target hardware comparison of YOLO26-seg models.

Model	Size (M Parameters)	Target Hardware
Nano	2.7	Mobile devices, Edge devices, CPU-only
Small	10.4	Laptop, low and medium tear GPUs
Medium	23.6	Server, high tear GPUs

.

Smaller variants provide faster inference and lower computational requirements, whereas larger variants generally achieve higher accuracy at the cost of increased latency. Consequently, the three variants were compared against each other.

Figure 12 compares the mask mAP50-95 obtained by the three YOLO26-seg variants during training. It a standard performance metric used in computer vision to evaluate object detection models and stands for mean Average Precision averaged across multiple Intersection over Union (IoU) thresholds, ranging from 0.50 to 0.95 with a step size of 0.05. It was selected because the downstream manipulation pipeline relies on segmentation masks to estimate component contours, centroids, orientations, and grasping points. The Small and Medium variants converge more rapidly during the initial training epochs, whereas the Nano model exhibits greater fluctuations and slower initial convergence. However, the differences progressively decrease, and all three models stabilize at comparable mask mAP50-95 values toward the end of training. Similar convergence trends were observed for mask precision and recall. These results indicate that increasing the model size did not provide a clear final segmentation-performance advantage for the considered dataset.

Table 4. Mask precision and recall achieved by the YOLO26-seg models at the final training epoch.

Model	Mask Precision	Mask Recall
Nano	0.9788	0.9690
Small	0.9628	0.9719
Medium	0.9597	0.9766

reports the mask precision and recall obtained by the three YOLO26-seg variants at the final training epoch. All models achieved similarly high values, with both metrics exceeding $0.95$. The Nano model obtained the highest mask precision, whereas the Medium variant achieved the highest mask recall. However, the Small model reached a precision comparable with the others. Overall, the differences among the three variants were limited and did not reveal a clear segmentation-performance advantage associated with increasing model size. These results support the selection of the Nano model, which provided comparable final segmentation performance with lower computational requirements and shorter inference times.

To assess deployment efficiency, inference time was measured on an additional set of 100 images acquired separately from the 250-image dataset. These images were not included in any of the dataset subsets and were used exclusively for timing evaluation under the same acquisition conditions and hardware configuration for all three model variants. Results are shown in Figure 13. The Nano model achieved the fastest and most consistent performance, with inference times centered around $0.4$ msec. The Small model approximately doubled the latency, reaching about $0.75$ msec while maintaining very low variability. In contrast, the Medium variant exhibited the highest computational cost, with a median inference time close to $2.0$ msec and a substantially larger spread. These results highlight the expected trade-off between model complexity and execution speed, with Nano providing the most suitable solution for real-time applications.

Considering both segmentation performance and computational efficiency, the YOLO26-seg Nano model was selected for deployment. Although the Nano variant showed slower convergence and greater variability during the initial training epochs, all three models reached comparable final values of mask mAP50-95, precision, and recall. Since the Small and Medium variants did not provide a clear improvement in final segmentation performance, their higher computational cost was not justified for the considered application. The Nano model therefore represented the most favourable trade-off between segmentation quality and inference efficiency.

4.2. Laboratory Experimental Tests and Discussion

The laboratory experimental campaign provided a preliminary assessment of the complete automated assembly cycle under the considered operating conditions. Its purpose was to verify the integration of the system modules, identify the main operational limitations, and collect initial observations on system behaviour under different illumination conditions. The test protocol consisted of multiple assembly cycles, each involving the sequential detection, picking, and placement of components according to a predefined bill of materials. The components were initially distributed randomly on the feeding surface to replicate industrial conditions, and the system was required to autonomously complete the assembly process. A luxmeter was used during the laboratory campaign to keep track of the changes in illumination and to relate the system behaviour to different lighting conditions. The experimental protocol considered a total of 30 assembly cycles, each corresponding to the assembly of one complete set of parts. These 30 cycles were intentionally distributed across three lighting conditions in order to evaluate the robustness of the system with respect to environmental illumination: 10 cycles were performed in low-light conditions, with no artificial lights and with natural light obscured; 10 cycles were carried out in medium-light conditions, with only natural light available; and 10 cycles were executed in high-light conditions, with both natural and artificial illumination. At the beginning of each cycle, the components were distributed randomly, reproducing the variability expected in a real feeding scenario. The operator was allowed to move some components only if potential obstructions were visible before the start of the automated sequence. A cycle was considered completed when the system determined that the assembly sequence had been fully executed. In order to classify the outcomes rigorously, three possible conditions were defined.

• A cycle was classified as autonomous success (ordinary operation) when the system completed the assembly autonomously, considered the result compliant with the specification, and the final assembled part was indeed compliant on inspection.
• A cycle was classified as an assisted success when the system recognized that it could not complete the assembly autonomously, requested human intervention, and then successfully finalized the process, still producing a compliant assembly.
• A cycle was classified as failure when the system completed the cycle autonomously and judged the result compliant, while the real assembly was actually non-compliant.

As reported in Figure 14, the experimental results demonstrate that the proposed robotic assembly cell achieved almost reliable operation, with 70% of the cycles completed without any human intervention and an additional 23% of the cycles resulted in assisted success. These results indicate a promising level of operational reliability, but they also show that the system is not yet fully robust under all tested conditions. In particular, the 7% failure rate represents a critical aspect, as these cases correspond to undetected non-compliant assemblies. We acknowledge that the limited experimental campaign does not provide statistically representative estimates for the long-term operation. Nevertheless, the tests were useful for identifying the main limitations of the current prototype and for distinguishing between assisted recoveries and undetected process failures. Therefore, a future more detailed analysis of both failures and assisted-success cases is planned to identify the main limitations of the current system and to distinguish between perception-related issues, manipulation errors, and fixture-related constraints.

Figure 15 reports the assembly outcomes as a function of the measured illuminance. Autonomous-success cycles were observed across the investigated range, from approximately 20 lux to nearly 1000 lux. Within this preliminary campaign, no failure was directly attributed to incorrect component classification. However, because only ten cycles were performed for each broad illumination condition and the measurements were not uniformly distributed across the illuminance range. Results therefore provide only an initial indication that illumination was not the dominant cause of the observed non-autonomous outcomes.

All failure cases were associated with manipulation-related issues and process monitoring. Specifically, failure occurred because the cylindrical spacer became stuck during insertion or due to unsuccessful grasping operations involving mainly washers and screws, resulting in assemblies being completed without the corresponding component. These events were not detected because the current grasp verification strategy relies on a relatively simple check of the gripper state, which proved insufficient in some situations. Furthermore, no in-process quality control is currently performed to verify the correct placement of components or the completeness of the assembly after each operation.

These findings indicate that overall the feasibility of the proposed automated assembly approach is confirmed. However, future developments should focus on strengthening grasp verification mechanisms and introducing intermediate quality inspection procedures throughout the assembly process. Such improvements would significantly reduce the likelihood of undetected assembly errors and further increase the reliability of the proposed system.

5. Conclusions

This work presents a comprehensive approach to the conceptualisation and implementation of a modular robotic cell specifically designed for the automated pre-assembly of heterogeneous metallic hardware, such as screws, washers, and spacers. The primary aim was to address the critical need for flexibility and reconfigurability in high-mix, low-volume production environments, which are typical of SMEs. By adopting an open architecture approach based on the ROS2 middleware, the research demonstrated how heterogeneous industrial components can be integrated into a cohesive, interoperable system. The study detailed a feasible technical pipeline, considering a tolerance chain analysis, the design of custom 3D-printed self-aligning fingertips, and the deployment of a YOLO-based image segmentation module for real-time component localization and pose estimation. In conclusion, the experimental results validate that the integrated ROS2-based system provides a promising foundation for autonomous industrial assembly, achieving a 70% success rate for fully autonomous cycles. A significant finding of this work is the effectiveness of the system’s modular software architecture and coordination module, which successfully identified operational limits in 23% of cases (assisted success), thereby requesting human intervention and preventing the production of non-compliant assemblies. The study also confirmed the effectiveness of the AI-driven vision pipeline; the comparison of YOLO26-seg variants indicated that the Nano model offers the most balanced solution for industrial deployment, providing high localization accuracy with a median inference time of just $0.4$ msec. Furthermore, the system demonstrated good environmental resilience in the analysed test scenarios, as experimental cycles performed reliably across a wide range of illumination conditions (20 to 1000 lux). Despite these successes, the research identified specific bottlenecks that remain as open challenges, such as the perception robustness for low-contrast components or the detection of mechanical jamming during assembly. Future work will therefore focus not only on improving grasp verification, trajectory planning, and in-process quality monitoring, but also on conducting a substantially larger experimental campaign. This campaign will include repeated tests under controlled illumination and operating conditions to quantify system performance and uncertainty through appropriate statistical analysis. Finally, this research provides a scalable architecture, that has the potential to improve cost-effectiveness when flexibility and repeated reconfiguration are economically relevant. This supports SMEs to transition from manual bottlenecks toward intelligent, reconfigurable automation.