A new automatic method for demoulding plastic parts using an intelligent robotic system

Industrial environments are usually surrounded by hazardous situations. One of them is the dangerous machinery, that can cause severe damage to human workers. For instance, in the toy industry, there are production ovens which are at high temperatures. The features of the handled material should be also considered as they can be composed of toxic substances that could harm humans likewise.

Although factories invest large amounts of resources in safety systems to prevent accidents through the integration of laser barriers and other measures and equipment, the traditional manufacturing processes have really high production targets with short cycle times which rise the stress of the operators. Human operators work for several hours on the same task; thus, they are more prone to get distracted and to make mistakes resulting in dangerous situations.

As mentioned, manual industrial processes are often repetitive and composed of tedious tasks and movements, especially in the toy manufacturing industry, which is the focus of this work. This process to create a plastic toy is based on the following tasks: first, operators fill the moulds with the plastic material, then they introduce the mould into a rotomoulding oven; they put the mould into an air cooler after that; finally, the operators place the mould (which is still at high temperatures) in the demoulding zone to gently extract the hot pieces that are inside the moulds. Operators repeat this process for several hours manually managing several moulds at the same time which increases their stress level.

This paper presents a new intelligent robotic system capable of performing the demoulding task of the entire toy manufacturing process by carrying out the most labour-intensive part of the process. Therefore, this approach directly reduces the stress and potential injuries to operators who can perform other dexterous and human-based tasks. This system is composed by the usual machinery of the toy manufacturing process (rotomoulding oven, air cooler, moulds, etc.), external devices such as RGB-D cameras, pneumatic actuators, emergency buttons, and safety laser scanners, and a UR10e collaborative robot arm. All mentioned devices and machinery have been integrated at INDUSTRIA AUXILIAR JUEMA S.L., an actual doll manufacturing company.

The main contribution of this work is the development of a new collaborative robotic system capable of performing a traditional manufacturing task such as the demoulding of soft plastic pieces autonomously, contributing to the creation of new future smart factories such as in [3]. As mentioned previously, during the extraction of the pieces, the mould is still at high temperatures, which means that the plastic pieces inside are deformable and flexible. Therefore, it completely changes the usual paradigm of rigid handling due to the possible deformations that pieces may suffer, so higher forces must be applied to accurately remove them from the mould in order to not damage or break them. Finally, with this approach, a reduction of the operator’s stress has been achieved, thus allowing them to perform other tasks with higher dexterity requirements or relocating them to the other sub-tasks of the process where the physical effort is less intensive.

The rest of the paper is structured as follows. In Sect. 2, different works with related objectives or developments are commented. In Sect. 3, it explains the steps of the process in the developed pipeline. Then, in Sect. 4, the experiments and results of the approach are shown to ensure the efficiency and accuracy of the system. Finally, in Sect. 5, some conclusions are described.

The state of the art related to the innovation and automation of industrial processes is really extended. The rise of the competitiveness between manufacturing factories has forced the integration of new technologies, especially in the robotic manipulation of soft and deformable materials. In the manufacturing processes, human operators manipulate different kinds of objects composed by soft material, such as plastic pieces, tyres, soles, and even food. One example of the automation of an industrial process handling deformable objects is [1], where authors present an innovative robotic system to manipulate clothes and automate the stitching process of cloth and a foam pad. However, they manipulate planar objects, which reduces the number of unknowns and uncertainty in comparison with 3D objects as proposed in this paper. In [10], the authors work with linear objects to perform the shape control without running real-time simulations or solving optimization problems, and they base their work on a partition of the nodal coordinates that allows deriving a control law directly from tangent stiffness matrices. Another approach about intelligent manipulation is proposed in [6], where authors present a new control framework for manipulating soft objects applying deep reinforcement learning to deform them and achieve a desired shape. Although the linear objects are more similar to this case, they are still simpler, and one of the main contributions of these works is based on simulation. Other kinds of learning examples are explained in [5], specifically learning from demonstration, a technique that creates policies from example state to action mappings. That approach defines a teacher, which provides the examples or demonstrations to the learner, according to the two different phases that the authors define to organize these techniques: the first one, the acquisition of the data, and the second one, the deriving of a policy from that information. These novel techniques could be useful for the demoulding task due to the complexity for the robot and the possibility of recording data from the expert operators. However, the gathering of information in our use case is really complex; operators use both hands applying several forces and carrying out determined trajectories to demould the pieces. In [4], the authors have implemented a novel approach called Advice-Operator Policy Improvement (A-OPI). Demonstration learning improves policies when the number of samples is larger; however, A-OPI synthesizes new data from a student execution and a teacher advice. Results show great potential as they surpass the performance of teleoperation. However, the system works for low-level motion control, which makes it impossible to integrate into our system due to the complexity of the task execution.

The approach presented in this paper is based on a high-effort required task, such as in [13], where it explains a strategy to improve the performance of current commercial industrial robots using a force-impedance controller. Nevertheless, our proposal’s aim is to create a system where operators and robots can work together, so these kinds of industrial robots are not allowed in a collaborative task. In [16], the authors present a force control loop used in a collaborative robot for sanding materials and state that collaborative robots can perform the same sanding task with similar results to industrial robots. The main difference with the work presented in this paper is the great dexterity of the robot movements needed when demoulding pieces instead of planar trajectories of that paper. In [18], it explains the current situation of robot applications in the food industry, which has high-quality requirements, and concludes with the enormous number of possible tasks that robots can carry out in this sector. Although the food sector is a more common research field than the toy one, similarities can be found with this industry, which has very strict quality requirements in the final product. Contrary to the other manufacturing fields, children’s doll production industry has not been much explored, and there are no robotic solutions integrated in real industrial environments, a fact that this work aims to change. In [17], it proposes a system that allows a safe cooperation between humans and high-payload robots during industrial tasks performance thanks to a tactile floor with spacial resolution able to define static safety zones or dynamic ones considering the current position of the joints and velocities. To integrate this system, companies need to implement major infrastructural changes in their factories. The solution presented in this paper aims to integrate two safety laser scanning devices to cover the entire working space and the areas shared between the operator and the robot in order to ensure human safety.

Plastic sector is really extensive and is composed by different kinds of tasks, such as in [14], where the authors developed a new approach which used the discrete properties of the moulds to create the configuration of the plastic injection moulds. The common features between this plastic material and the children’s dolls highlight the challenges associated with handling soft materials, and how their physical behaviour varies based on temperature. The study of these properties is tedious and complicated as shown in [11], where the authors present an automatic shape control of deformable wires based on the unknown deformation model or the mechanical properties of the object. They just use some visual features to calculate the deformation and improve the manipulation. A related work is described in [8], where the authors report the first autonomous robotic solution for the USB wires soldering task. The similarities with the approach presented in this paper arise due to the repetitive and closed-cycle nature of the demoulding process.

As mentioned previously, toy industry is a manufacturer and traditional sector which is composed mainly by small-and-medium enterprises (SMEs), which are not usually able to develop or integrate new technologies into their production processes. In [19], it explains the possibility of automation in SMEs with a collaborative robot and learning-based vision systems. They propose an automatic system for object detection and quality control of products, together with a multi-functional gripper capable of performing different operations without tool changing. In [7], the authors present a study about the reinforcement learning in contact manipulation robotic tasks which explains the increase of the research in this field. In addition, they explain different kinds of contact manipulation tasks depending on the dynamic interaction of the objects. Our use case could be classified in the ‘pushing tasks’ group, where authors expose the difficulty from the unknown and nonlinear dynamics when the robot must perform gentle manipulation on objects of different sizes, shapes, and textures. In [12], the authors present a review about the smart robotic manufacturing representing the new epoch of a higher degree of intelligence in industrial tasks carried out by the robots. In [15], the authors present a user-friendly model of robot skills, tested in real industrial scenarios, specifically designed for operators with no prior knowledge of robotics. This aspect holds significant value for this work because one of the goals of this approach is to prevent operator replacement and, instead, to involve them in the operational process by making it more accessible and manageable. In addition to the task automation, the approach supports and trains the operators to make the transition to the robotizing of the demoulding task easier. This fact could be supported with the proposal explained in [2], where the authors present a demonstration learning framework for robots wherein they developed a force-based acquisition system to capture the task essence in two distinct scenarios: human-human and human-robot collaboration. The purpose is to extract task-specific features and transfer these skills to the robot for the purpose of extracting task-specific features and transfer these skills to the robot. In this context, the involvement of the human factor becomes crucial as it plays a vital role in extracting task-related features, facilitating the robot’s learning process, and fostering collaboration between the human and the robot during task execution. They tested the system in the co-manipulation of objects and assembly of simple interlocking parts. In contrast to deformable objects, rigid manipulation makes the collaboration and the acquisition of task parameters easier. Nevertheless, due to the complexity of the soft object parameters and models, in this work, the authors are able to extract the knowledge of the operators to programme the robot trajectories for the demoulding task. To improve the manipulation of flexible objects, in [9], it presents a survey of model-based off-line manipulation planning systems. An initial object classification is stated, dividing into volumetric, planar, and linear, then they categorize the planning strategies according to the type of goal: path planning, folding/unfolding, topology modifications, and assembly. Although model-based strategies are really useful to manipulate deformable objects, in the focused scenario of this paper, the deformation parameters may change during the task performance due to the temperature variation.

As stated in this section, currently, there is no approach as presented in this paper: a new intelligent robotic system capable to perform the demoulding of soft plastic pieces in an autonomous way. In addition, this innovative system can collaborate with the human operator in order to finish the entire toy manufacturing process.

The demoulding task of the toy manufacturing industry is one of the several steps of the production process. The company has divided into different groups the operators in order to cover all steps; one group produces the liquid plastic material from raw powder and some additives, and another group uses this liquid material to manufacture the solid pieces relying in the rotomoulding manufacturing process. Then, while one group paints, sews the hair, and puts the eyes on the head, the other group assembles the pieces. Finally, the assembled doll is packaged and distributed to retailers.

In this section, the results obtained in the automatic system of the demoulding task are shown in order to measure the success of the tasks performed by the developed system. The process has been split into three different sub-tasks: detection, grasping, and demoulding. As this is a sequential process, the output of the perception module is the input of the grasping module, and its output is the input of the demoulding module. In this way, failures are just considered in the current module and not in the next ones allowing us to measure the success and accuracy of the different modules of the system. For the detection and grasping module, own developers of the system checked at each cycle if the output is correct or not. However, the demoulding module is checked by the expert operator of the factory, due to their experience and knowledge of the final product.

In order to calculate the following production and quality metrics, 10 cycles of each mould have been carried out. A cycle considers the process from the time the robot removes the parts from the mould until the mould is returned to the demoulding area after the rotational moulding process.

As mentioned previously, the system will perform the task in 10 complete cycles. Furthermore, in order to compare the system performance, the robot will carry out the task at 100%, 90%, and 80% of the maximum force (225 Newtons) and velocity (1000 mm/s) allowed for the robot. Next, tables (Tables 1, 2, and 3) show the results of the different tests carried out and the outcomes of the metrics. After 10 cycles, depending on the kind of piece/mould, the maximum number of pieces could be 20 (heads and bodies) or 40 (arms and legs).

Next, some conclusions are exposed from the results obtained in the experiments. The arm and leg parts are more difficult to demould than the other pieces due to their elongated shape, which makes the material more concentrated and the piece more compact. This fact forces the system to carry out a two-step demoulding procedure for these pieces. Firstly, it is extracted the upper part of the piece in order to enable the airflow between the inner surface of the mould and the piece, thereby facilitating the extraction process. Afterwards, it is demoulded in the lower segment. Moreover, as the body part is the largest part, the robot needs to make more movements during demoulding, which means that it takes more time than the other parts. Finally, the headpiece is the easiest to demould due to its uniform and rounded shape. The headpiece also holds a large amount of air inside it, and the gripper fingers can apply vacuum and make the demoulding task easier. Given the diverse shapes of each piece, the robot follows varying paths during the demoulding process. Elongated pieces need supplementary movements to initially release the upper segment before demoulding the rest of the piece. Furthermore, the body and head are demoulded directly.

In general terms, it is possible to see a forward correlation between the percentage of force and velocity applied, and the APS (Eq. 1) which justifies that the demoulding task requires high effort by the operators. However, the APS of the detection module has no significant changes as it is independent of the velocity or force of the robot; it just depends on the camera. Finally, the grasping module changes because the part is soft and flexible, so the robot spends too much time from detecting the parts until it grasps them; the parts may deform due to the lack of temperature.

As mentioned previously, APS (Eq. 1) metric measures the success of each module independently, and it is possible to see how it increases as force and velocity increase. However, the detection module is independent to the robot’s performance, so it remains roughly constant. ATP (Eq. 2) metric has the same behaviour as the previous one as it is directly correlated with the force and velocity of the robot’s performance. Meanwhile, the robot at 80% spends 9 s to demould 1 head, increasing 10% the time is reduced by 1.5 s, and finally, at 100% is 1 s faster still. In industrial processes in which machinery is working 24 h, there is a production of 9600 heads at 80% in a day; in comparison with 13,292 heads at 100%, there is a huge difference for the company. ATPO (Eq. 3) and NTPH (Eq. 4) metrics are similar in the three cases as the time of the rotomoulding process and the cooling is included and they are constant; however, in 1 year period, the difference between the robot at 80% and 100% represents more than 1000 produced pieces. Finally, MOPH (Eq. 4) considers the APS (Eq. 1), and a huge difference is shown in the table between the 3 demoulding tests. From 80 to 90%, there is an 75.82% increase in performance, and from 90 to 100%, a 34.83%. These increments represent a remarkable production improvement in the long term. In the period of 1 year, the robot at 80% runs 9855 operations, and each operation is a complete cycle of each mould, which means 118,260 produced pieces, at 90% the robot produces 207,927 pieces, and at 100%, it produces 280,355 pieces. The difference of 20% is more than double the number of pieces produced in a year. In summary, conducting tests at different capacity levels provides a holistic understanding of a robot’s performance, safety margins, efficiency, result validity, and adaptability. This approach ensures that the robot’s capabilities are thoroughly examined and validated for both optimal performance and safe operation, while also accounting for potential changes in operational conditions over time.

Finally, Fig. 14 illustrates the trajectories traced by the robot during the demoulding process of the headpiece. This visual representation offers a valuable perspective on the movements and displacements the robot has performed in its environment. The 3D graph provides a better understanding of the relationship between the position of the robot and the force applied by the robotic tool. The blue line represents the position of the robot along the task path, and the red circles represent the amount of force obtained from the robotic tool (the larger the size, the more force).

As shown in Fig. 14, once the detection algorithm provides the piece extraction point, the robot moves to the mould and performs the trajectories to grip the piece and demould it. The trajectory begins at the ‘Init” marker. Subsequently, the robot moves forward to point ‘A”, which designates the detected extraction point of the headpiece. Following this, the robot executes a rotation movement applying force at point ‘B’, where it grasps the piece using the gripper. Finally, the robot moves from ‘B’ to ‘C’ and performs an upward pull to successfully extract the piece from the mould. Larger circles appear when the robot grasps, rotates, and pulls up to remove the piece. However, the amount of force provided by the robot is only the one detected in the end effector, not in all the joints. For a better understanding of the task performance data, Fig. 15 shows the end effector positions and orientations of the robot, as well as the forces during the demoulding task. These values are shown along an iteration axis.

Upon comparing both previous plots (Figs. 14 and 15), the relationship among points ‘A’, ‘B’, and ‘C’ becomes apparent. The robot initiates movement moving to point ‘A’, where the orientation and position stabilize at constant values; then, the position undergoes a brief alteration while the force demonstrates an upward trend. Subsequently, the robot pivots and advances with force application upon reaching point ‘B’, coinciding with the Z-axis rotation matching that of the initial point. Lastly, as the robot reaches point ‘C’, it pulls upward, revealing both positional and force elevation along the Z-axis. Summarizing, while extracting the piece from the mould, the end effector records a force that surpasses 40 N. This implies that the other joints are also subjected to higher force magnitudes. Consequently, it is evident that the robot must use its full capacity, even operating near its limits, to successfully execute the demoulding process.

In conclusion, this work proposes a new automatic and collaborative robotic system able to perform the demoulding task of soft plastic pieces in a real industrial environment of a doll manufacturing company. This system is based on a vision algorithm to detect the moulds and the extraction point of the pieces to send them to the robot and perform the demoulding. The force requirement of this task is a challenge for a collaborative robot; however, by executing similar trajectories to the expert operators, it is possible to carry out the task correctly.

Finally, this collaborative task allows operators to perform other tasks that require more dexterity while the robot performs the task, avoiding injuries. The results show us the success rate of the system, which makes it useful for the company, provides an increase in production, and serves as the first step for the technological transference for small companies.