Camera arrangement optimization for workspace monitoring in human-robot collaboration

Citation: Ošˇcádal, P.; Kot, T.; Spurný, T.; Suder, J.; Vocetka, M.; Dobeš, L.; Bobovský, Z. Camera Arrangement Optimization for Workspace Monitoring in Human–Robot Collaboration. Sensors 2023,23, 295. https://doi.org/10.3390/s23010295 Academic Editor: Mariusz Kostrzewski Received: 13 October 2022 Revised: 23 December 2022 Accepted: 26 December 2022 Published: 27 December 2022 Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). sensors Article Camera Arrangement Optimization for Workspace Monitoring in Human–Robot Collaboration Petr Ošˇcádal 1,* , Tomáš Kot 1, Tomáš Spurný1, JiˇríSuder 1, Michal Vocetka 1, Libor Dobeš 2 and Zdenko Bobovský1,* 1 Department of Robotics, Faculty of Mechanical Engineering, VSB-TU Ostrava, 70833 Ostrava, Czech Republic 2MoravskoslezskýAutomobilovýKlastr, z.s., Business Incubator VŠB-TU Ostrava, 70833 Ostrava, Czech Republic *Correspondence: petr[email protected] (P.O.); [email protected] (Z.B.) Abstract: Human–robot interaction is becoming an integral part of practice. There is a greater emphasis on safety in workplaces where a robot may bump into a worker. In practice, there are solutions that control the robot based on the potential energy in a collision or a robot re-planning the straight-line trajectory. However, a sensor system must be designed to detect obstacles across the human–robot shared workspace. So far, there is no procedure that engineers can follow in practice to deploy sensors ideally. We come up with the idea of classifying the space as an importance index, which determines what part of the workspace sensors should sense to ensure ideal obstacle sensing. Then, the ideal camera positions can be automatically found according to this classified map. Based on the experiment, the coverage of the important volume by the calculated camera position in the workspace was found to be on average 37% greater compared to a camera placed intuitively by test subjects. Using two cameras at the workplace, the calculated positions were 27% more effective than the subjects’ camera positions. Furthermore, for three cameras, the calculated positions were 13% better than the subjects’ camera positions, with a total coverage of more than 99% of the classified map. Keywords: workspace monitoring; camera; human–robot interaction; collaboration; sensors network 1. Introduction Human–robot collaboration (HRC) is currently one of the most developing areas of research, with applications not only in industrial and service robotics but also in elderly care, rescue robotics, intelligent vehicles and aircraft, rehabilitation technology and space applications [ 1 , 2 ]. HRC combines human capabilities with robot precision and efficiency [ 3 ]. Robots that enable collaboration in a shared space with humans are referred to as collaborative robots [ 4 ]. Collaborative robot workplaces are designed such that the robot performs the part of the operation that requires high precision or that may be non-ergonomic, repetitive, or even dangerous for humans, while the human does the part of the cycle that requires dexterity, intuition, or unique decision making [ 5 ]. Such collaboration cannot be performed without a shared workspace [3]. When sharing a workspace, collisions can occur when individual pieces of hardware collide with each other or with some objects in the environment. If only robots move in the shared workspace, their work cycle can be uniquely determined by the control system and then collisions can be avoided by applying suitable algorithms, such as elastic strips [ 6 ], artificial potentials [ 7 ], or other similar variants [ 8 , 9 ]. In the case where the robot’s workspace is shared with a human, the system cannot unambiguously define the worker’s motion, but only predict it to a limited extent, and, thus, collisions between the robot and operator may occur during the work cycle [ 10 , 11 ]. In HRC systems, operator safety is the most important criterion [ 12 – 14 ]. In order to ensure maximum operator safety, various safety-related requirements are imposed on the collaborative workstation according to Sensors 2023,23, 295. https://doi.org/10.3390/s23010295 https://www.mdpi.com/journal/sensors Sensors 2023,23, 295 2 of 18 standards, see [ 15 , 16 ] for a review. According to [ 17 , 18 ], for example, safety zones are defined in the workspace, in which the magnitudes of robot speeds or actuator torques are adjusted to a safe value. Despite the fact that it pays to invest in smart factories [ 19 ], workplaces with robots still do not sufficiently solve some problems related to sharing workspace with humans. A collision, even a safe one, most often leads to an immediate stop of the robot [ 14 ], or other collision resolution strategies are used based on the collected data [ 20 , 21 ]. In all cases, however, this means increased duty cycle time [ 12 ], resulting in higher energy consumption and making the operation more expensive. For this reason, it is advantageous to use non-contact systems for collision prediction and avoidance [ 22 , 23 ]. Monitoring the shared space is often based on camera systems to detect humans [ 24 ] or any obstacles [ 25 ]. One use of these camera systems is in conjunction with human–robot interaction (HRI) systems that, for example, use special gloves [ 26 , 27 ] to inform the operator of a possible collision with a robot. This allows the operator to react to this event and adapt his or her movements to avoid a collision. Another way of avoiding a predicted collision is to re-plan the robot’s trajectory in time according to the collected real-time data [12,28]. A disadvantage of camera systems is the possibility of obscuring their surveillance area [ 29 ]. The robot’s movement and the operator’s activity can obscure this space during the operation. This creates a volume in the workspace that is not monitored; thus, it cannot be clearly determined whether or not an obstacle is present, which may lead to a collision. Eliminating or minimising the shaded volume of the workspace can be achieved by increasing the number of cameras and by choosing their appropriate placement [25]. The correct position of cameras has a significant impact on the performance of the system, and this position depends on the purpose of the system itself. Although camera positions can have a large impact on the quality of the acquired data, there are currently not many methods for determining the optimal number of cameras and their placement for workspace monitoring. For example, the work in [ 30 ] describes a method for determining the positions of cameras to provide the best input information relative to the actions performed by a human (recognize motion, learn activities, take measurements, etc.). The authors in [ 31 ] describe a simple 2D algorithm that minimises camera view frusta overlap, while not considering any shadowing caused by obstacles, nor different importance of various sections of the workplace. The paper [ 32 ] investigates how to deploy the cameras in such a way that the 3D data error is minimized. First, an analytical uncertainty method based on minimizing the error criterion was used, followed by evolutionary optimization methods similar to genetic algorithms. There is also a method for placing sensors in 3D space called CamHunt [ 33 ], described in the example of using it to place cameras inside rooms and the whole building. CamHunt uses a 3D grid partitioning of the environment for camera placement, where the goal is that each voxel is seen by at least one camera. However, it is not described in considering the usage of the robot and its influence. On the other hand, it uses the methodology of multi-camera system placement focused on human presence [34]. Another method addresses placing two different types of sensors (depth and presence) in a shared workspace [ 35 ]. It proposes the placement of sensors according to a probabilistic framework computed based on the presence or depth map of obstacles on their image plane, including the robot manipulator maps. This paper discusses the classification of the workspace shared between the robot and human. Based on the classified map, the camera covering the most space with respect to the importance of the space is searched. Multiple cameras can be found this way: each additional camera added to the workspace covers the space not covered by the cameras already used. This process also takes into account dynamic objects (e.g., the robot). The main motivation is to achieve the necessary workspace coverage using as few cameras as possible, in order to minimize purchase costs, maintenance costs and energy consumption. Our hypothesis is that camera(s) arranged by human subjects using intuition will provide less coverage of the workspace compared to camera(s) arranged by the algorithm. Sensors 2023,23, 295 3 of 18 2. Space Classification In order to design camera positions for workplace monitoring, the workplace must be thoroughly studied. The idea of space classification evaluates parts of the space using an importance index. The importance index describes the impact of the examined volume in the workspace. For clarity, the importance index will be visualised using a colour gradient (see Figure 1), where pure red represents the most important area (an area that is very important for monitoring as there is a very high risk of collision with the robot), green represents the areas with a lower importance (less important for monitoring), and white represents areas, where no collisions can occur (zero importance) and, thus, there is no need for any monitoring. Figure 1. Color coding of space classification importance index ranking from unimportant (white) to very important (red). The first step is to discretise the workspace into a voxel grid. Each voxel in the grid (a small cube) represents an area with a specific value of the importance index which will be calculated by classification of the space. Dangerous objects, such as the robot, will be integrated into the grid. To classify the space, classification functions have to be defined. These functions represent the importance around the hazardous elements in the voxel map and calculate the above-mentioned importance indices for all voxels. The functions are defined according to the type of workplace, technology, size, and the level of importance wanted to be ensured in the workplace. The number of classification functions is variable. Only the two most important will be described here: classification based on position of hazardous objects in the voxel map, and classification based on direction of movement of hazardous objects. Based on the type or technology of the workplace, it is also possible to classify, for example, by radiation (heat, light, etc.), maximum energy induced by impact, dangerous tools or manipulated objects (e.g., the possibility of a cut wound caused by sharp sheet metal), etc. The position classification represents and stores information about accumulated proximity of each voxel from all voxels containing dangerous objects (obstacles) in the voxel map. For a specific voxel with centre point pxyz , the position classification value kd xyz can be calculated as kd xyz = n ∑ i=1Ld max−    pxyz −oi    (1) where     pxyz −oi     represents the distance between the centre of the voxel under investigation pxyz and the centre of a voxel oi containing a dangerous obstacle, and Ld max represents the chosen threshold value for distance. Voxels with     pxyz −oi    ≥Ld max are not included at all, their contribution is considered zero rather than negative as the equation would suggest. This classification creates an imaginary volume around obstacles, whose importance decreases with the distance from the obstacles. Figure 2demonstrates this principle on a simplified image in 2D space for better clarity. The voxels are represented by squares; the black square contains a dangerous obstacle. Colour coding is according to Figure 1. The second classification focuses on the velocity and direction of movement of the hazardous obstacles. It also captures the space around the particles, but voxels in the direction of motion are classified as more important than particles in the opposite direction. For a specific voxel with centre point pxyz , the velocity classification value kv xyz can be calculated as kv xyz = n ∑ i=1   ||v|| 2  1+pxyz −oi·v     pxyz −oi    ·    v      ·Lv max−    pxyz −oi     (2) Sensors 2023,23, 295 4 of 18 Figure 2. Simplified 2D visualization of position classification of voxels around an obstacle (the black voxel). The impact of the direction of motion is captured by the dot product between the motion vector v and the vector pointing from the dangerous voxel oi to the investigated voxel pxyz . The dot product is multiplied by a distance factor, decreasing the resulting value with increasing distance from the obstacle. Again, this is performed only for voxels with     pxyz −oi    <Ld max . A simplified 2D example of velocity classification is show in Figure 3. Figure 3. Simplified 2D visualization of velocity classification of voxels around an obstacle (the black voxel); the arrow represents the movement vector of the obstacle. It is also important to take into account the access of the worker to the monitored workspace. Typically, the human operator approaches the work area from a single direction or a few directions (e.g., from the front and the left side). This can be simply described by access planes. In the vicinity of these planes, the most frequent occurrence of the operator is assumed and, thus, the importance of collision checking is increased. This parameter can be a key feature of classification functions. It determines the importance (weight) of the classification functions by applying a scaling factor fw xyz calculated for a specific voxel and naccess planes as follows: fw xyz = n ∑ i=1 Li max −di xyz Li max (3) where Li max represents the maximum distance from the i-th access plane to be considered and di xyz represents the distance of the voxel from the i-th plane. The importance value fw xyz is equal to one directly at the access plane and decreases linearly until reaching zero at the chosen threshold distance Li max . For all voxels with di xyz >Li max , we consider fw xyz = 0. The visualisation in Figure 4demonstrates in 2D the weight value corresponding to one access plane (the left side of the image). The worker access plane is a simple way to include a worker in the classification map. If there are some known predetermined movements that the worker performs in the workplace, the plane can be replaced or supplemented by discrete points or bounding volumes (boxes) in which the worker is frequently located, and weights of adjacent voxels can be affected based on the distances from these points. Sensors 2023,23, 295 5 of 18 Figure 4. Visualisation of the importance weight in a 2D workspace with one access plane (depicted by the black arrows). The total classification index for a particular voxel is calculated as the sum of the values of all chosen classification functions calculated for this voxel, each multiplied by an optional weight factor wi . These additional weights can be introduced to fine-tune the relative importance of individual classification functions, as needed by the specific use-case. The resulting sum is then multiplied by the overall importance scaling factor fw xyz given by the access plane(s). If the number of classification functions is n, the total classification index for a given voxel is calculated as follows: kxyz =fw xyz· n ∑ i=1 wi·ki xyz (4) In our case, we consider two classification functions—see Equations (1) and (2)—and, thus, this equation can be written specifically as follows (the visualisation can be seen in Figure 5): kxyz =fw xyz wdkd xyz +wvkv xyz(5) Figure 5. Simplified 2D visualization of the total classification index of voxels around an obstacle (the black voxel). 3. Camera Classification Camera classification index is a value that describes how good a camera with a particular position and orientation is in monitoring the workplace. The camera classification cis calculated as the sum of classification values kxyz of all voxels visible for the camera (k0xyz): c= n ∑ i=0 k0i(6) Sensors 2023,23, 295 6 of 18 To find the voxels that are visible to the camera, we cast a ray from the camera origin through each individual pixel of the camera image. Every voxel intersected by some ray is considered visible; see an example in Figure 6, where visible voxels are drawn as white squares and invisible as grey squares. Grey squares represent parts of the workplace that cannot be monitored by the camera. If there is an obstacle (e.g., the robot) in the workspace, the voxels containing the obstacle (black squares in Figure 6) block the rays and, thus, the voxels behind the obstacle are also considered invisible. Figure 6. Using camera rays (red lines) to detect visible voxels in the grid; white voxels are visible, grey voxels are invisible, black voxels represent an obstacle. The camera classification index (6) represents the coverage of the grid by the camera with respect to the importance of the space. If two cameras capture the same number of voxels, their classification indices may vary according to the importance of the voxels captured. Since the obstacles can move in the environment during the work cycle (typically a robot arm following a trajectory in a cycle), the camera coverage varies during time; see an example in Figure 7. Figure 7. Camera coverage of a voxel map with a moving obstacle (black voxels) during three time moments; (a) t1; (b) t2; (c) t3. In this case, the total coverage is calculated as the sum of the camera coverage cat each time step during the whole movement cycle from t= 0 to t=T. The percentage of camera coverage over time ( crel ) is the sum of the coverage c(t) at each time step divided by the sum of all voxel classification values (including voxels not visible to the camera) during the whole cycle period: crel =∑T t=0c(t) ∑T t=0∑n i=0ki(t)·100 [%](7) 4. Finding the Optimal Camera Location To find the best camera placement for workspace monitoring, we can use Equation (7) to evaluate the cameras and pick the one with the highest value. This can be undertaken Sensors 2023,23, 295 7 of 18 either by an optimization algorithm, or by simply trying all possible camera locations (in discrete distances, to obtain a finite number of possible cameras). There are six values to optimize: three coordinates for camera position and three values for camera orientation in 3D space. To speed up the optimization process, it is possible to limit the number of variables to three if we determine the camera orientation explicitly by some function based on the camera location. The camera orientation significantly influences the classification; if the orientation is inappropriately chosen (for example, away from the direction of the voxel grid), the camera may not sense the space at all. To ensure maximum camera coverage, each camera can be focused on the imaginary centre of gravity of the workspace voxel map (the camera axis passes through the centre of gravity). The centre of gravity of the voxel grid is calculated as the centre of gravity of a system of mass points, where the points are defined as voxel centres, and the classification function value represents their mass. Figure 8shows the centre of gravity for the simplified example. Figure 8. Example of a center of gravity (drawn as a white dot) of a simplified 2D voxel grid. If multiple cameras are to be used to cover the space, the best camera (with the highest coverage crel ) is evaluated first, and then all voxels covered by this camera are devalued by setting their classification index kxyz to zero (separately for each time step during the work cycle T). Then, a new grid centre of gravity is computed for the remaining voxels (not covered by the previous camera), and the best camera for these now conditions is found again. This way, a new camera is selected, which serves as a complementary camera for the previous one. The whole algorithm of designing a camera subsystem is summarized by the flowchart in Figure 9. The process starts by generating the space classification voxel map according to Section 2, where each voxel has an importance index value that describes its importance for monitoring. This first step depends on the description of the workspace layout, the robot task and on the chosen definition of classification functions; see Equations (1)–(5). Then, the centre of gravity (focus point) is calculated for the voxel map. The positions of possible cameras are then generated in the defined available space, each camera is oriented towards the focus point. For all these potential camera positions, the camera classification is calculated according to Section 3, which gives each camera its relative coverage; see Equations (6) and (7). The position with the largest coverage is then selected as the best position for that iteration and this camera is added to the list of proposed cameras. If the total coverage ( cT ) achieved by all proposed cameras (or the single camera, if we are in the first iteration) is sufficient, the process is terminated. If more coverage is needed, it is checked whether the maximum possible number of cameras has not been reached already (the limit must be specified by the user). If this limit has been reached, the process ends, since the required coverage could not be achieved. If the limit has not been exceeded, all voxels visible from the recently added camera are effectively removed from the classification map by setting their importance value to zero and the next iteration of the process is started to find another complementary camera to the already calculated cameras. The result is then a list of camera positions and orientations. Sensors 2023,23, 295 8 of 18 Figure 9. Flowchart describing the whole algorithm. 5. Experiment An experiment was designed to demonstrate the effectiveness of our system. The experiment involves a workstation with the UR3 robot performing an assembly task; see Figure 10. The robot is responsible for picking up a screw from the feeder at station C and screwing it at assembly stations A and B, one at a time (the sequence is C-A-C-B, repeated in a cycle). Meanwhile, the human operator replaces the parts at stations A and B; the operator changes the part at station A while the robot operates at station B and vice versa. The positions of the screwdriver tool in the robot arm in each station are shown in Figure 11. The robot trajectory for the task cycle is visualised in Figure 12a. Visualisation of the overall workspace volume is shown in Figure 12b in the form of a voxel grid with a grid size of 5 cm. Figure 12c shows the operator access plane. Sensors 2023,23, 295 9 of 18 Figure 10. Experimental workplace with the UR3 robot; letters A and B denote the assembly stations, letter C denotes the supply feeder. Figure 11. Target positions of the UR3 manipulator during the work cycle: ( a ) assembly station A; (b) assembly station B; (c) supply feeder at station C. Figure 12. Simulation of the experimental workplace: ( a ) robot trajectory between stations A, B, and C; ( b ) voxelgrid coveringthe workspacevolume (voxel size is 5cm); ( c )operatoraccessplanetotheworkplace. Sensors 2023,23, 295 16 of 18 the centre of gravity of the classified voxel map (where the classification index determines the weight of the position) is computed to determine the camera focus point. For the generated camera positions with an orientation towards the centre of gravity of the classified voxel map, the camera indices are calculated to represent the coverage of the classified voxel map (the sum of the classification indices or the voxels that are visible for the camera). The camera with the largest index covers the most space with respect to the space importance indices. This whole methodology was verified in a real experiment, where engineers familiar with the workplace task were asked to design a camera subsystem to provide security in the workplace. Subjects were tasked to place one, two, and then three cameras to monitor the workplace. These results were compared with the calculated camera positions. The results confirmed an expected trend that more space is covered when more cameras are used. However, the calculated cameras clearly covered more space than the manually placed cameras. While the average coverage achieved by the subjects was 55.0% for one camera, 72% for two cameras, and 87% for three cameras, the coverage of the calculated cameras was 91.9% for one camera, 98.8% for two cameras, and 99.8% for three cameras in the workspace. Furthermore, the system managed to achieve a better workplace coverage with just a single camera compared to what 16 out of 30 testing subjects managed with even three cameras. Based on the results, we can confirm the hypothesis that the camera system designed by engineers’ intuition achieves lower space coverage compared to cameras arranged by the mathematical model proposed by us. The purpose of the paper is to provide a methodology for designing a camera subsystem, that will be able to monitor the important sections of a shared workplace in order to prevent collisions between a human worker and a collaborative robot. The importance is, thus, influenced primarily by proximity to the robot arm (the proposed cameras will be able to monitor the robot and its surroundings). The paper does not solve the question how to actually detect the human worker in the monitored workspace using the cameras, that is out of the scope of this research. Future work can focus, for example, on the impact of various classification functions and the recommended choice of weights. Another topic worth investigating is the application of optimisation algorithms or neural networks to find the best camera position and orientation instead of the brute-force grid method. With an optimisation algorithm or a neural network, the calculation could be faster, more parameters could be found than just the camera location, and camera positions could be found more precisely than what the grid discretisation is capable of. Furthermore, multiple cameras could be searched together rather than finding one camera and then a complementary second camera. Author Contributions: Conceptualisation, P.O., Z.B. and T.K.; methodology, P.O., Z.B. and T.K.; software, P.O. and T.K.; validation, T.S. and Z.B.; formal analysis, T.K., T.S. and M.V.; investigation, J.S., M.V. and L.D.; resources, T.S. and J.S.; data curation, P.O. and T.S.; writing—original draft preparation, P.O., T.K. and T.S.; writing—review and editing, T.K. and T.S.; visualisation, T.K. and P.O.; supervision, Z.B.; project administration, Z.B., funding acquisition, L.D. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by the Research Platform focused on Industry 4.0 and Robotics in Ostrava Agglomeration project, project number CZ.02.1.01/0.0/0.0/17_049/0008425 within the Operational Programme Research, Development and Education. This article has also been supported by the Specific Research Project SP2022/67 and financed by the state budget of the Czech Republic. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: There are no data available to share. Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results. Sensors 2023,23, 295 17 of 18 References 1. Bauer, A.; Wollherr, D.; Buss, M. Human-Robot Collaboration: A Survey. Int. J. Hum. Robot. 2008,5, 47–66. [CrossRef] 2. Feleke, A.G.; Bi, L.; Fei, W. EMG-based 3D hand motor intention prediction for information transfer from human to robot. Sensors 2021,21, 1316. [CrossRef] [PubMed] 3. Wang, L.; Gao, R.; Váncza, J.; Krüger, J.; Wang, X.V.; Makris, S.; Chryssolouris, G. Symbiotic Human-robot collaborative assembly. CIRP Ann. 2019,68, 701–726. [CrossRef] 4. Semeraro, F.; Griffiths, A.; Cangelosi, A. Human–robot collaboration and machine learning: A systematic review of recent research. Robot. Comput. Integr. Manuf. 2023,79, 102432. [CrossRef] 5. Vysocky, A.L.; Novak, P.E. Human-Robot collaboration in industry. MM Sci. J. 2016,9, 903–906. [CrossRef] 6. Brock, O.; Khatib, O. Elastic strips: A framework for motion generation in human environments. Int. J. Robot. Res. 2002 ,21, 1031–1052. [CrossRef] 7. Warren, C.W. Global path planning using artificial potential fields. In Proceedings of the 1989 International Conference on Robotics and Automation, Scottsdale, AZ, USA, 14–19 May 1989; IEEE: Piscataway, NJ, USA; Volume 1, pp. 316–321. [CrossRef] 8. Raibail, M.; Rahman, A.H.A.; AL-Anizy, G.J.; Nasrudin, M.F.; Nadzir, M.S.M.; Noraini, N.M.R.; Yee, T.S. Decentralized MultiRobot Collision Avoidance: A Systematic Review from 2015 to 2021. Symmetry 2022,14, 610. [CrossRef] 9. Kot, T.; Bobovský, Z.; Brandstötter, M.; Krys, V.; Virgala, I.; Novák, P. Finding Optimal Manipulator Arm Shapes to Avoid Collisions in a Static Environment. Appl. Sci. 2021,11, 64. [CrossRef] 10. Tang, K.-H.; Ho, C.-F.; Mehlich, J.; Chen, S.-T. Assessment of handover prediction models in estimation of cycle times for manual assembly tasks in a human–robot collaborative environment. Appl. Sci. 2020,10, 556. [CrossRef] 11. Mainprice, J.; Berenson, D. Human-robot collaborative manipulation planning using early prediction of human motion. In Proceedings of the 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, Tokyo, Japan, 3–7 November 2013; pp. 299–306. 12. Liu, H.; Wang, L. Collision-free human-robot collaboration based on context awareness. Robot. Comput. Integr. Manuf. 2021 ,67, 101997. [CrossRef] 13. Liu, H.; Wang, Y.; Ji, W.; Wang, L. A Context-Aware Safety System for Human-Robot Collaboration. In Procedia Manufacturing; Elsevier BV: Amsterdam, The Netherlands, 2018; Volume 17, pp. 238–245. [CrossRef] 14. Haddadin, S.; De Luca, A.; Albu-Schäffer, A. Robot Collisions: A Survey on Detection, Isolation, and Identification. IEEE Trans. Robot. 2017,33, 1292–1312. [CrossRef] 15. Fryman, J.; Matthias, B. Safety of Industrial Robots: From Conventional to Collaborative Applications. In Proceedings of the ROBOTIK 2012; 7th German Conference on Robotics, Munich, Germany, 21–22 May 2012; IEEE: Piscataway, NJ, USA, 2012. 16. Rosenstrauch, M.J.; Krüger, J. Safe human-robot-collaboration-introduction and experiment using ISO/TS 15066. In Proceedings of the 2017 3rd International Conference on Control, Automation and Robotics (ICCAR), Nagoya, Japan, 24–26 April 2017; IEEE: Piscataway, NJ, USA, 2017; pp. 740–744. [CrossRef] 17. Mansfeld, N.; Hamad, M.; Becker, M.; Marin, A.; Haddadin, S. Safety Map: A Unified Representation for Biomechanics Impact Data and Robot Instantaneous Dynamic Properties. IEEE Robot. Autom. Lett. 2018,3, 1880–1887. [CrossRef] 18. Zanchettin, A.M.; Ceriani, N.; Rocco, P.; Ding, H.; Matthias, B. Safety in human-robot collaborative manufacturing environments: Metrics and control. IEEE Trans. Autom. Sci. Eng. 2016,13, 882–893. [CrossRef] 19. Kostrzewski, M.; Sompolski, K.; Królikowski, T. How profitable is it to apply Industry 4.0 solutions in blister factories? In Procedia Computer Science; Elsevier BV: Amsterdam, The Netherlands, 2022; Volume 207, pp. 3281–3289. [CrossRef] 20. Parusel, S.; Haddadin, S.; Albu-Schäffer, A. Modular state-based behavior control for safe human-robot interaction: A lightweight control architecture for a lightweight robot. In Proceedings of the 2011 IEEE International Conference on Robotics and Automation, Shanghai, China, 9–13 May 2011; IEEE: Piscataway, NJ, USA, 2011; pp. 4298–4305. [CrossRef] 21. Haddadin, S.; Albu-Schaffer, A.; De Luca, A.; Hirzinger, G. Collision Detection and Reaction: A Contribution to Safe Physical Human-Robot Interaction. In Proceedings of the 2008 IEEE/RSJ International Conference on Intelligent Robots and Systems, Nice, France, 22–26 September 2008; IEEE: Piscataway, NJ, USA, 2008; pp. 3356–3363. [CrossRef] 22. Hermann, A.; Mauch, F.; Fischnaller, K.; Klemm, S.; Roennau, A.; Dillmann, R. Anticipate your surroundings: Predictive collision detection between dynamic obstacles and planned robot trajectories on the GPU. In Proceedings of the 2015 European Conference on Mobile Robots (ECMR), Lincoln, UK, 2–4 September 2015; pp. 1–8. 23. Kot, T.; Wierbica, R.; Ošˇcádal, P.; Spurný, T.; Bobovský, Z. Using Elastic Bands for Collision Avoidance in Collaborative Robotics. IEEE Access 2022. [CrossRef] 24. Li, G.; Liu, Z.; Cai, L.; Yan, J. Standing-Posture Recognition in Human–Robot Collaboration Based on Deep Learning and the Dempster–Shafer Evidence Theory. Sensors 2020,20, 1158. [CrossRef] [PubMed] 25. Ošˇcádal, P.; Spurný, T.; Kot, T.; Grushko, S.; Suder, J.; Heczko, D.; Novák, P.; Bobovský, Z. Distributed Camera Subsystem for Obstacle Detection. Sensors 2022,22, 4588. [CrossRef] [PubMed] 26. Grushko, S.; Vysocký, A.; Ošˇcádal, P.; Vocetka, M.; Novák, P.; Bobovský, Z. Improved Mutual Understanding for Human-Robot Collaboration: Combining Human-Aware Motion Planning with Haptic Feedback Devices for Communicating Planned Trajectory. Sensors 2021,21, 3673. [CrossRef] 27. Grushko, S.; Vysocký, A.; Heczko, D.; Bobovský, Z. Intuitive Spatial Tactile Feedback for Better Awareness about Robot Trajectory during Human–Robot Collaboration. Sensors 2021,21, 5748. [CrossRef] Sensors 2023,23, 295 18 of 18 28. Wang, L.; Schmidt, B.; Nee, A.Y. Vision-guided active collision avoidance for human-robot collaborations. Manuf. Lett. 2013 ,1, 5–8. [CrossRef] 29. Maric, B.; Jurican, F.; Orsag, M.; Kovacic, Z. Vision based collision detection for a safe collaborative industrial manipulator. In Proceedings of the 2021 IEEE International Conference on Intelligence and Safety for Robotics (ISR), Tokoname, Japan, 4–6 March 2021; IEEE: Piscataway, NJ, USA, 2021; pp. 334–337. [CrossRef] 30. Bodor, R.; Drenner, A.; Schrater, P.; Papanikolopoulos, N. Optimal Camera Placement for Automated Surveillance Tasks. J. Intell. Robot. Syst. 2007,50, 257–295. [CrossRef] 31. Gonzalez-Barbosa, J.; Garcia-Ramirez, T.; Salas, J.; Hurtado-Ramos, J.; Rico-Jimenez, J. Optimal camera placement for total coverage. In Proceedings of the 2009 IEEE International Conference on Robotics and Automation, Kobe, Japan, 12–17 May 2009; IEEE: Piscataway, NJ, USA, 2009; pp. 844–848. [CrossRef] 32. Olague, G.; Mohr, R. Optimal camera placement to obtain accurate 3D point positions. In Proceedings of the Fourteenth International Conference on Pattern Recognition (Cat. No.98EX170), Brisbane, QLD, Australia, 20 August 1998; IEEE: Piscataway, NJ, USA. [CrossRef] 33. Becker, E.; Guerra-Filho, G.; Makedon, F. Automatic sensor placement in a 3D volume. In Proceedings of the 2nd International Conference on PErvsive Technologies Related to Assistive Environments–PETRA ’09, Corfu, Greece, 9–13 June 2009; ACM Press: New York, NY, USA, 2009. [CrossRef] 34. David, P.; Idasiak, V.; Kratz, F. A Sensor Placement Approach for the Monitoring of Indoor Scenes. In Smart Sensing and Context; Kortuem, G., Finney, J., Lea, R., Sundramoorthy, V., Eds.; EuroSSC 2007; Lecture Notes in Computer Science; Springer: Berlin/Heidelberg, Germany, 2007; Volume 4793. [CrossRef] 35. Flacco, F.; De Luca, A. Multiple depth/presence sensors: Integration and optimal placement for human/robot coexistence. In Proceedings of the 2010 IEEE International Conference on Robotics and Automation, Anchorage, AK, USA, 3–7 May 2010; IEEE: Piscataway, NJ, USA; pp. 3916–3923. [CrossRef] 36. Ošˇcádal, P.; Heczko, D.; Vysocký, A.; Mlotek, J.; Novák, P.; Virgala, I.; Sukop, M.; Bobovský, Z. Improved Pose Estimation of Aruco Tags Using a Novel 3D Placement Strategy. Sensors 2020,20, 4825. [CrossRef] [PubMed] 37. Huczala, D.; Ošˇcádal, P.; Spurný, T.; Vysocký, A.; Vocetka, M.; Bobovský, Z. Camera-Based Method for Identification of the Layout of a Robotic Workcell. Appl. Sci. 2020,10, 7679. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.