Enhancing Eficiency and Safety Through Autonomous Navigation in the Intensive Care Unit

Abstract

The use of autonomous systems in Intensive Care Units (ICUs) has become incredibly important, especially during the COVID-19 pandemic. This period has overwhelmed both ICUs and hospitals, halting many other medical activities and causing significant challenges. This project aims to develop a navigation system tailored specifically for the ICU environment, adapting it to the unique procedures and regulations of that setting. Due to the critical conditions of ICU patients, strict rules dictate precise requirements for navigation, necessitating a context-specific approach. This work will propose a comprehensive navigation system capable of safely guiding from point A to point B within an ICU while addressing the critical issues present in such environments. Unlike traditional Nav2 systems, it will feature specialized collision avoidance components designed specifically for ICU settings, taking into account both contextual demands and the chosen approach. This will involve implementing a multilayered protection technique and employing active movements to prevent collisions with dynamic obstacles.

Keywords

• eol>Autonomous navigation
• Collision avoidance
• Artificial Intelligence
• Reinforcement Learning

1. Introduction

addressing navigation challenges. Not all tasks can be addressed solely using ROS components, particularly in Robotic systems play a vital role in dangerous environ- contexts where specific constraints are required, necesments and risky situations where human presence is sitating the creation of custom nodes. One of the most limited. The COVID-19 pandemic has emphasized the ne- critical issues in ICU navigation is collision avoidance for cessity for hospitals to adopt automated robotic solutions, dynamic objects. While objects within the intensive care particularly in treating patients with highly contagious unit are typically stationary, situations may arise where diseases. These systems minimize the risk of infection for healthcare personnel urgently need to reach a patient’s medical staf, addressing the challenges encountered dur- bedside during emergencies. Therefore, it’s essential for ing the early stages of the pandemic. This project aims to the robot to navigate while avoiding collisions and withdevelop a navigation system specifically designed for in- out impeding the path of doctors who need to reach the tensive care units (ICUs). It will generate dynamic paths, patient promptly. adhere to contextual rules, avoid collisions with swiftly In addressing navigation challenges within intensive moving individuals, and efectively manage potentially care unit (ICU) environments, researchers have proposed critical situations. several approaches outlined in the literature. One such

The challenge of navigation within confined envi- method involves the utilization of potential cost maps ronments is commonly addressed and resolved using to navigate around crowded areas [ 47, 33, 54, 34 ]). This established solutions like ROS (Robot Operating Sys- strategy allows robots to map out regions of high congestem) [49, 32] and Nav2 [ 48, 35 ]. ROS, introduced in tion and adjust their paths accordingly to avoid potential 2007 through collaborative eforts among universities collisions. Another critical aspect is collision avoidance and robotics-focused companies, has become a standard for dynamic objects within the ICU. Given the unpreframework for navigation and robotic applications. It dictable nature of medical emergencies, it is essential ofers modular components and user-friendly monitor- for robotic systems to adaptively maneuver around moving and simulation tools, making it the standard solution ing obstacles to ensure the safety of both patients and for navigation tasks. The extensive contributions and medical personnel. Lastly, researchers have explored the vast library of modular components associated with ROS concept of employing critical profiles to modulate robotic and Nav2 have solidified their status as essential tools for behavior based on the level of urgency or criticality encountered in the ICU environment, as proposed in the work by [50]. This approach enables robots to dynamically adjust their navigation strategies in response to

3. Related Works

Based on the requirements specified in the previous section, it is possible to identify 2 macro-areas of study which, although strongly related to each other, are 2 topics addressed by previous works independently; they are collision avoidance of dynamic objects and dynamic object identification.

3.1. Collision Avoidance in Navigation

integrating these diverse approaches, robotic navigation systems can enhance safety and eficiency within ICU settings, ultimately contributing to improved patient care outcomes [ 9, 38, 8 ].

These supplementary nodes enable robots to navigate safely through swiftly moving humans. The navigation system can leverage the same framework as proposed by Nav2 but with added enhancements. In navigation systems, route planning is facilitated through a cost map that assigns a cost to each point in the environment, aiding in obstacle identification and the generation of the most eficient trajectory. By integrating a cost map derived from high-trafic areas, it becomes feasible to chart routes that circumvent locations frequented by medical personnel. This approach enhances navigation eficiency and minimizes the likelihood of interference with healthcare professionals.

The issue of avoiding collisions while navigating in con

ifned spaces is extensively studied from various angles.

Dealing with stationary objects has been thoroughly explored, and numerous algorithms ready for use are available in robotic systems like ROS. However, navigating around moving objects remains a challenge due to its close ties with specific contexts and constraints. Unlike stationary objects, dynamic ones demand a diferent ap2. ICU Specific Problems proach. Their unpredictable nature necessitates swift responses to prevent collisions, and predicting their fuIn designing a navigation system, it’s crucial to outline ture positions in the near term could prove invaluable for all requirements, constraints, and assumptions pertinent anticipating and avoiding potential collisions. Multiagent to the context, particularly within an ICU environment. collision avoidance during navigation can be seen as a An Intensive Care Unit (ICU) represents an environment cooperative task where each agent plays a role in avoidof utmost criticality, as all hospitalized individuals are ing collisions [ 42, 36 ]. In environments like ICU (Indoor facing severe clinical conditions posing a real and signifi- Closed Environments), communication between agents cant threat to their lives. Consequently, ICUs are meticu- is often absent, with sensor data being the only source lously structured to maximize the operational eficiency of information about positions and velocities [ 10, 1, 40 ]. of medical personnel. Beds and associated equipment Understanding this task involves considering two main are strategically positioned within the room, typically approaches: trajectory-based and reaction-based methin proximity to essential utilities such as oxygen or UPS ods [ 41 ]. The reaction-based approach focuses on shortoutlets, as well as room alarms. Additional items like term responses, while the trajectory-based method plans chairs or desks cannot be accommodated, and access for over a longer duration. While the latter yields smoother patients’ relatives is restricted, with staf expected to trajectories, it can be computationally intensive. Despite complete their tasks promptly and vacate the room im- this classification, there’s potential in combining these mediately. This protocol gained heightened importance approaches. By leveraging Reinforcement Learning (RL), during the COVID-19 pandemic, as prolonged presence computationally expensive operations can be moved to in the room increased the risk of infection. ofline training stages, with an online policy eficiently

In designing the navigation system, it’s imperative to queried during runtime. This hybrid approach ofers ensure that robot movements prioritize human safety advantages from both methods. Everett et al. propose while adhering to the following constraints: Medical per- an RL-based solution aimed at determining velocity vecsonnel trajectories must remain unaltered; the robot must tors to reach goal positions swiftly while avoiding collinever impede their movement. Obstacles’ movements sions with other agents [ 42, 17 ]. Their method employs are unpredictable, requiring reactive adjustments to de- an LSTM network capable of handling various numbers tected trajectories. In critical situations where no action of agents. Interestingly, the LSTM is used not in a secan resolve potential collision, the robot must halt im- quential time-based manner, but to encode each agent mediately, allowing humans to evade collision. Upon the sequentially, utilizing the final hidden layer for subsecessation of a dangerous event or potential collision, the quent steps. This design ensures consistent dimensions robot must revert to its initial goal, either by returning regardless of the number of agents involved. Another to the initial path, reaching the destination, or devising a crucial consideration is the impracticality of applying new path. Navigation interruption or updating the final the trajectory-based method in scenarios involving hugoal point should be feasible. Lastly implementing a max- mans, as predicting human paths is challenging due to imum speed limit for the robot in all situations ensures unpredictable needs or rules. It’s often unclear beforesafety. hand where a human might need to go or if they may need to return to their initial position for various reasons. to identify static obstacles within a local map, which is Therefore, collision avoidance systems become valuable then integrated into the navigation system. Dynamic only when agents are near the target robot and can be obstacles require a diferent approach, as their velocities analyzed over a short period, during which their move- necessitate fast and specific algorithms capable of calcuments can be predicted to some extent. Otherwise, the lating both position and velocity. A method proposed data becomes too noisy for efective trajectory planning. by In [ 43 ] was introduced as a technique for identifyWhile RL has been utilized in various studies, it’s essen- ing dynamic objects using spatiotemporal norms derived tial to specify constraints on the behavior of other agents from points gathered by robotic sensors, such as Lidar to yield meaningful results, as highlighted by Everett. or depth cameras. This method involves clustering or Generic approaches where each agent exhibits arbitrary creating point clouds from the sensor-detected points, behavior may not lead to efective collision avoidance representing potential objects for analysis. Subsequently, strategies. An alternative approach is presented in [53], a spatiotemporal norm analysis is applied to these point which leverages Lidar data to identify dynamic objects, clouds to determine if the objects are undergoing transtrack their movements, and employ the ORCA algorithm lation or rotation. This approach enables the recovery [52] to compute collision-free paths. This method of- of an object’s position using sensor data. By analyzfers another avenue for efective collision avoidance in ing the covariance (a measure of sparsity) of the point dynamic environments. cloud, possible rotations and the object’s radius can be

Optimal Reciprocal Collision Avoidance (ORCA) [52] reconstructed, providing valuable information for collistands out as a velocity-based planning technique sion avoidance systems. It’s crucial to note that these renowned for its ability to ensure collision avoidance results are based solely on points identified by the Lidar in both static and dynamic environments, boasting high or depth camera. This means that the central position and scalability. ORCA evaluates the velocities of all agents covariance calculations are limited to the visible parts involved and delineates cones representing potential col- of the objects; any points hidden from the sensors are lision scenarios. Subsequently, through an optimization not considered in the analysis. Therefore, this analysis is process, it determines the minimum velocity necessary inherently related to the visible portions of the objects. to navigate out of these collision cones.

While ORCA ofers robust collision avoidance capabilities, it operates on an optimization search paradigm, 4. ICU Navigation System which implies that it can find a solution multiple times if one exists. However, it does have a couple of drawbacks.

Firstly, it assumes homogeneity among objects within the workspace, whether static or following the same navigation policy. Deviating from this assumption can lead to catastrophic outcomes, as noted by Vince Kurtz [ 44 ]. Secondly, its reliance on optimization introduces a notable delay, potentially resulting in missed solutions. Such delays could pose challenges in environments requiring swift decision-making.

In response to these limitations, was proposed an alternative approach. This method involves predicting the motion of dynamic objects over a short timeframe using an LSTM RNN with online training, as suggested in [ 41 ]. Once these predictions are available, they are integrated into a Nonlinear Probabilistic Velocity Obstacles algorithm. This adapted algorithm efectively handles collision avoidance in static environments and accounts for objects moving along predictable trajectories, derived from short-term predictions.

Navigating an ICU has specific demands that must be

satisfied in a robust architecture that is modified and extended for that. The presented navigation system is based on the Nav2 infrastructure in the ROS2 environment with custom components (or nodes) specific to the ICU context. This project has a Gazebo simulation with rViz monitoring. The architecture comprises a standard navigation system supplemented by three additional nodes designed to enhance collision avoidance capabilities.

The proposed system navigates the robot from point A to point B using a cost map generated by SLAM [ 2 ]. In an Intensive Care Unit (ICU), where minimizing instrument presence and their impact on medical staf movement is crucial, the map’s variability is assumed to be rare. It’s created extemporaneously and maintained throughout the robot’s operational lifecycle, with manual updates only in exceptional cases. This approach doesn’t significantly constrain the project, as static objects not on the map can still be identified and managed by local and collision avoidance components with regularly updated perspectives.

Given Nav2’s adeptness at handling navigation in closed environments, this project focuses on collision avoidance, the primary ICU navigation requirement. Avoiding static or dynamic objects and medical personnel is achieved through a multilayered approach, each

3.2. Dynamic Object Identification

In navigating around obstacles, whether they are static or dynamic, the first step is to detect them using sensors installed on the robot. Ready-to-use components, such as those found in ROS and Nav2, are commonly used

4.2. Potential Areas • Localmap: Identifies new objects via local maps As per the requirements, the navigation system must priand adjusts navigation accordingly, efective for oritize avoiding collisions with medical personnel. Hence, static but less so for dynamic obstacles. it’s vital to identify areas with a significant probability • Potential Areas: Modifies the cost map based on of encountering people or obstacles and steer clear of the likelihood of encountering obstacles in cer- trajectories leading into those areas. In the cost map, potain areas, thereby reducing the probability of tential values range from integers 0 to 255. High values, choosing paths through them. like 255 and 254 (considered as lethal costs), indicate a • Collision Monitor: A ROS2 node that reduces the high probability of collision with an object, while very robot’s speed when obstacles are nearby. low values such as 0 or 1 (representing free space costs) • Emergency Guard: Adjusts movement profiles or denote no obstacles and safe navigation. Each value sigcapabilities based on the distance from obstacles, nifies a distinct collision probability based on proximity potentially slowing down or stopping the robot. to obstacles. The component identifies contact points of lidar rays with objects (referred to as hit points ). For • Dynamic Collision Avoidance: Alters the robot’s every point in the hit point set, it’s inferred that an object trajectory to avoid collisions dynamically. exists at that point, thus necessitating an increase in its Each layer addresses specific scenarios, bolstering safety potential. Conversely, if there’s no hit (a point not in and meeting ICU requirements efectively. ), the potential is decreased (indicating the absence of an object at that point) following an exponential func4.1. Localmap tion with decay parameter . Figure 1 illustrates how potential values evolve (depicted in the black box in the During navigation, sensor data create a cost map close upper right corner) as an object moves, and how potento the robot with a high-frequency update. This updated tials decrease over time. Adjusting enables control over and high-quality map is used to adapt the trajectory, the decay rate. avoiding obstacles and keeping the original track (defined by the Planner Server) as much as possible. This node is (, + ∆ ) = min(MAX_PTNL; (1) already present in the Nav2 framework and is used as is, (, ) * − Δ + ( ∈ )) for this reason will not be analyzed anymore.

In equation 1, the potential is refreshed every ∆ . If point is hit by a ray (in ), its value increases; oth

erwise, it decreases gradually over time with a decay pe- predicts the new minimum distance that will potentially riod. To maintain flexibility in navigation, the potential is be detected. capped ( _ ), allowing trajectories through those areas albeit with a low probability. Subsequently, ˆ() = () + (() − ( − 1)) (2) the potential map is integrated into the environment map and regularly updated with new potential values (refer to Figure 2).

If the predicted distance falls below a warning threshold ( _ ), the navigation system applies a reducedspeed movement profile. However, if it drops below a critical threshold ( _), the robot transitions into a blocking profile. This blocking profile indicates a highly critical situation, prompting the robot to halt, allowing humans to intervene and avoid a collision.

⎧ ⎪ ⎨

⎪⎩ ˆ() < _; ˆ() < _ ; = ℎ

(3) Once the predicted distance exceeds the critical threshold, the robot remains in a critical state until both the warning threshold and normal conditions are met. For safety reasons, the robot maintains its critical state until it returns to a normal operating state. This behavior is depicted in the state diagram shown in Figure 3.

4.3. Collision Monitor

The Collision Monitor, a node available within the Nav2

framework, serves as a crucial safety feature in the navigation system. Positioned just before the command is dispatched to the robot, its primary function is to ensure safe navigation. When the robot approaches an obstacle, the Collision Monitor intervenes by reducing its velocity. Specifically, it may decrease the velocity to a fraction of the original command, such as 20%, to prevent collisions. However, if the robot is not close to any obstacles, the command remains unchanged. Since the Collision Monitor is an integral part of the Nav2 framework and is utilized without modification, it will not be further analyzed as its functionality is standardized and deemed suficient for the system’s safety requirements.

4.4. Emergency Guard

Hit Point Identification : Identifying contact points

Emergency Guard is a custom ROS2 node that improves of laser rays with surrounding objects. Each point the capability of a standard collision monitor. The com- represents a hit point on an object. ponent identifies the minimum distance with a generic object (both static and dynamic), evaluates the variation Environment Filtering: Filtering out all hit points origicompared to the previous detection, and based on that nating from the ICU environment. Since this is focused

A crucial aspect of collision avoidance is the identifi

cation of dynamic objects and the extraction of their positions and velocities. To achieve this, an approach inspired by the work of Raphael Falque [ 43 ] has been adapted to suit the specific context. The algorithm encompasses the following steps: on dynamic object detection, static objects within the environment are disregarded. Filtering is accomplished using a kNN (k-Nearest Neighbors) algorithm. A kNN model is trained using the environment’s cost map. For each hit point, nearby points are identified using the Nearest Neighbors of the trained model. If at least one point is at a critical level (indicating the presence of an environmental object), the hit point is filtered out. This ifltering process helps eliminate noise from detection. the movement (see Figure 4). If the area exceeds a predeifned threshold, dynamic behavior is identified.

Point Cloud Creation: After filtering, the remaining hit points are clustered using a DBScan algorithm. This approach ofers the advantage of dividing points into Figure 4: Images depicting the Naive approach’s reconstrucgroups without needing to specify the number of clusters tion of the movement area over time, with dynamic behavior beforehand. However, it requires careful tuning to ensure identified based on the threshold-exceeded area the creation of sparse clusters. This tuning typically involves setting a high value for the epsilon parameter (ples)paanrdama elotewr v(alue_forthem).inTimheu mresnuulmtinbgercol ufsstaemrs- = 1 ∑︁ ( − )( − ) (5) form the point clouds representing the identified objects. || || ∈ To eliminate isolated points resulting from incorrect distance measurements, clusters with a small number of When analyzing the variation of covariance, two elements are filtered out. Additionally, any points not scenarios may arise: rotation of the object or a reduction associated with a cluster are removed from consideration. in distance between the object and the robot. While theoretically, significant variability in covariance over some time could indicate movement, strong measurement noise can lead to frequent changes in covariance even without real movement. Therefore, covariance analysis is not reliable for evaluating and identifying object dynamics.

Dinamicity Identification : Each point cloud is analyzed to determine its movement characteristics. Initially, the center of mass for the points in point cloud , denoted as , is computed. This is achieved by calculating the average position of its hit points, as shown in equation 4. Subsequently, the covariance of the point cloud, represented by equation 5, is determined. This covariance Dynamic Object Measurement: Given the list of dyprovides information about the dispersion of the points namic objects, their positions (or central positions of around the center of mass, aiding in recognizing the ob- points in their cloud) and velocities are recovered from ject’s movement pattern. previous points. This information is then utilized by the collision avoidance engine. Calculating velocity by ana = 1 ∑︁ (4) lyzing close positions in time, such as from two consecu|| || ∈ tive measurements, often results in significant positional errors and speed oscillations. To mitigate this, a period (or a number of measurements or odometry messages) is considered to calculate velocity (see equation 6). This approach helps reduce measurement errors and ensures smoother velocity estimation.

Detecting movement involves assessing significant variations in the center of the point cloud. However, noise in measurements can lead to fluctuations in the center that need to be filtered out. One common approach to identify variations or trends in sequential values is to use linear regression. However, this method is not suitable in this context because rapid generation of points results in a dataset with values that are nearly constant or exhibit very small variations. Consequently, the linear regression yields a zero coeficient with a prediction constant equal to the mean value. Any deviations from this fixed value are interpreted as errors rather than meaningful changes. To address this issue and mitigate the efects of noise and the high number of values, a Naive approach is proposed. This approach reconstructs the area of movement over time, defined as the rectangle encompassing

The Collision Avoidance node is designed to prevent col

lisions with dynamically moving objects that may pose an imminent threat to the robot. Its primary objective is not to define the entire navigation path towards the goal, but rather to handle complex dynamic situations efectively. When a critical situation arises, indicated by

5. Implementation

This project was developed on Ubuntu 22.04 using

ROS2 Humble distribution and Nav2. Custom nodes were implemented using Python 3.10 and interfaced with ROS services. Simulations and monitoring were conducted using Gazebo and rViz. The robot and environment were adapted from turtlebot3, with a custom world tailored to resemble an ICU environment (refer to Figure 5). objects being closer than a specific threshold, the node retrieves positional and velocity information of dynamic objects from the dedicated Dynamic Object Identification node (see section 4.5), along with the robot’s position and the goal pose. It then applies an algorithm to determine the appropriate speed to resolve the critical event. Once the critical situation is resolved (i.e., objects are no longer too close), the node is deactivated, and standard navigation by the Nav2 Plan Server resumes. This node is based 6. Results on the Deep Reinforcement Learning strategy proposed by Michael Everett [ 41 ]. The algorithm adopts a multi- The development process was marked by a series of thoragent approach capable of handling a variable number of ough tests and evaluations, both for individual custom dynamic objects. To accommodate the variable number nodes and for the entire system. This approach ensured of obstacles, they are standardized using an LSTM net- that each node could complete its task, ensuring safe bework. Each obstacle’s state is fed into an LSTM network, havior in the robot’s movement. In particular, detailed and only the last hidden layer is utilized in subsequent tests were conducted to explore a range of scenarios, inprocessing (see Figure 6). This condensed representation cluding edge cases, to identify and address any issues. of dynamic objects, along with the robot’s state, is used One of the main challenges encountered during developto generate a vector passed through two fully connected ment was the presence of noise in measurements, particlayers, resulting in a probability distribution for possible ularly evident in the contact points detected by sensors. actions. These fluctuations in position could compromise the system’s reliability, necessitating careful calibration and implementation of advanced filtering algorithms. The goal was to minimize the impact of noise and ensure proper interpretation of data by the system.

Another significant challenge was managing computational resources, critical for the proper functioning of the system, especially in real-time environments like robotics.

Insuficient resources could lead to delays in message processing and calculations, potentially afecting overall system performance. Consequently, optimizing resource fFoirgtuhreea6c:tFiounllsnetwork to generate a probability distribution usage through the implementation of parallelization techniques, optimization algorithms, or potential hardware upgrades was essential.

Ultimately, successfully addressing these challenges was crucial for the project’s progress. Through an integrated approach involving thorough testing, algorithm optimization, and resource management, the system was able to ensure reliable behavior in real-world applications.

7. Conclusion

The proposed solution adopts the architectural infras

tructure of a widely used navigation system based on Nav2, with customized extensions tailored specifically for the ICU environment. This standardized approach facilitated rapid prototyping and integration within the existing navigation framework.

Custom navigation components designed for the ICU environment contribute to enhanced safety characteristics during navigation. For instance, the Emergency Guard component complements the Collision Monitor by introducing diferent profiles that gradually restrict the robot’s movement freedom. These profiles adjust parameters such as speed or radius, crucial for collision avoidance algorithms, while also incorporating simple prediction mechanisms to preemptively address potential safety risks.

Potential maps play an important role in shaping the costmap based on the likelihood of encountering obstacles, thus enabling the generation of safer trajectories. Careful configuration of potential values, including setting appropriate decay periods, ensures the creation of up-to-date potential maps that accurately reflect recent object detections. Fine-tuning these parameters is essential to strike a balance between maintaining high potential values in areas with recent obstacles while avoiding excessive averaging that could diminish the efectiveness of this feature.

The collision avoidance algorithm prioritizes rapid decision-making through the adoption of an online reinforcement learning (RL) model. This node selectively activates only during critical situations, swiftly deactivating once the threat has passed. Importantly, in scenarios where the collision avoidance movement brings the robot into proximity with static objects, other nodes such as the Collision Monitor or Emergency Guard are triggered to avert collisions and uphold overall safety.

In summary, the integrated system ofers a robust solution for navigating within an ICU environment, ensuring a high level of safety through the coordinated eforts of multiple nodes. By addressing a range of potentially critical events with distinct functionalities, the system efectively meets the initial safety requirements of ICU navigation.

References

[1] Y. F. Chen , M. Liu, M. Everett , J. P. How , Decentralized non-communicating multiagent collision avoidance with deep reinforcement learning , in: 2017 IEEE International Conference on Robotics and Automation (ICRA) , 2017 , pp. 285 - 292 . doi: 10 . 1109/ICRA. 2017 . 7989037 .

[2] H. Durrant-Whyte , T. Bailey , Simultaneous localization and mapping: part i , IEEE Robotics & Automation Magazine 13 ( 2006 ) 99 - 110 . doi: 10 .1109/MRA. 2006 . 1638022 .

[3] G. Capizzi , S. Coco , G. L. Sciuto , C. Napoli , A new iterative fir filter design approach using a gaussian approximation , IEEE Signal Processing Letters 25 ( 2018 ) 1615 - 1619 . doi: 10 .1109/LSP. 2018 . 2866926 .

[4] M. Woźniak , D. Połap , C. Napoli , E. Tramontana, Graphic object feature extraction system based on cuckoo search algorithm , Expert Systems with Applications 66 ( 2016 ) 20 - 31 . doi: 10 .1016/j.eswa. 2016 . 08 .068.

[5] C. Napoli , G. Pappalardo , E. Tramontana , Z. Marszalek , D. Polap , M. Wozniak , Simplified firefly algorithm for 2d image key-points search , 2014 . doi: 10 .1109/CIHLI. 2014 . 7013395 .

[6] C. Napoli , G. Pappalardo , E. Tramontana, Improving files availability for bittorrent using a difusion model , 2014 , pp. 191 - 196 . doi: 10 .1109/WETICE. 2014 . 65 .

[7] G. Capizzi , F. Bonanno , C. Napoli , Hybrid neural networks architectures for soc and voltage prediction of new generation batteries storage , 2011 , pp. 341 - 344 . doi: 10 .1109/ICCEP. 2011 . 6036301 .

[8] M. Wozniak , C. Napoli , E. Tramontana, G. Capizzi, G. Lo Sciuto , R. K. Nowicki , J. T. Starczewski , A multiscale image compressor with rbfnn and discrete wavelet decomposition , volume 2015 -September , 2015 . doi: 10 .1109/IJCNN. 2015 . 7280461 .

[9] M. Woźniak , D. Połap , R. K. Nowicki , C. Napoli , G. Pappalardo , E. Tramontana, Novel approach toward medical signals classifier , volume 2015 -September , 2015 . doi: 10 .1109/IJCNN. 2015 . 7280556 .

[10] D. Połap , M. Woźniak , C. Napoli , E. Tramontana, R. Damaševičius , Is the colony of ants able to recognize graphic objects? , Communications in Computer and Information Science 538 ( 2015 ) 376 - 387 . doi: 10 .1007/978-3- 319 -24770-0_ 33 .

[11] M. Woźniak , D. Połap , M. Gabryel , R. K. Nowicki , C. Napoli , E. Tramontana, Can we process 2d images using artificial bee colony? , volume 9119 , 2015 , pp. 660 - 671 . doi: 10 .1007/ 978-3- 319 -19324-3_ 59 .

[12] B. A. Nowak , R. K. Nowicki , M. Woźniak , C. Napoli, Multi-class nearest neighbour classifier for incom- 978-3- 319 -03524-6_ 45 . plete data handling , volume 9119 , 2015 , pp. 469 - [ 23] G. Capizzi , G. L. Sciuto , C. Napoli , M. Woźniak, 480 . doi: 10 .1007/978-3- 319 -19324-3_ 42 . G. Susi, A spiking neural network-based long-term

[13] C. Napoli , G. Pappalardo , E. Tramontana, R. K. Now- prediction system for biogas production, Neural icki , J. T. Starczewski, M. Woźniak , Toward work Networks 129 ( 2020 ) 271 - 279 . doi: 10 .1016/j. groups classification based on probabilistic neural neunet . 2020 . 06 .001. network approach, volume 9119 , 2015 , pp. 79 - 89 . [24] C. Napoli , G. Pappalardo , E. Tramontana, Using doi: 10 .1007/978-3- 319 -19324- 3_8. modularity metrics to assist move method refac-

[14] G. Borowik , M. Woźniak , A. Fornaia , R. Giunta, toring of large systems , 2013 , pp. 529 - 534 . C. Napoli, G. Pappalardo , E. Tramontana , A soft- doi:10.1109/CISIS. 2013 . 96 . ware architecture assisting workflow executions [25] F. Bonanno , G. Capizzi, G. L. Sciuto , C. Napoli, on cloud resources, International Journal of Elec- G. Pappalardo, E. Tramontana , A cascade neutronics and Telecommunications 61 ( 2015 ) 17 - 23 . ral network architecture investigating surface plasdoi: 10.1515/eletel-2015-0002. mon polaritons propagation for thin metals in

[15] F. Bonanno , G. Capizzi, G. L. Sciuto , C. Napoli, openmp 8467 LNAI ( 2014 ) 22 - 33 . doi: 10 .1007/ Wavelet recurrent neural network with semi- 978-3-319-07173-2_3 . parametric input data preprocessing for micro- wind [26] F. Bonanno , G. Capizzi , A. Gagliano , C. Napoli , Oppower forecasting in integrated generation systems, timal management of various renewable energy 2015 , pp. 602 - 609 . doi: 10 .1109/ICCEP. 2015 . sources by a new forecasting method , 2012 , pp. 934 7177554 . - 940 . doi: 10 .1109/SPEEDAM. 2012 . 6264603 .

[16] G. Capizzi , G. L. Sciuto , C. Napoli , E. Tramontana , A [27] G. Capizzi , G. L. Sciuto , C. Napoli , R. Shikler, multithread nested neural network architecture to M. Wozniak, Optimizing the organic solar cell model surface plasmon polaritons propagation, Mi- manufacturing process by means of afm measurecromachines 7 ( 2016 ). doi:10.3390/mi7070110. ments and neural networks, Energies 11 ( 2018 ).

[17] D. Połap , M. Woźniak , C. Napoli , E. Tramontana, doi:10.3390/en11051221. Real-time cloud-based game management system [28] G. Capizzi , C. Napoli , L. Paternò , An innovavia cuckoo search algorithm , International Journal tive hybrid neuro-wavelet method for reconstrucof Electronics and Telecommunications 61 ( 2015 ) tion of missing data in astronomical photometric 333 - 338 . doi: 10 .1515/eletel-2015 -0043. surveys 7267 LNAI ( 2012 ) 21 - 29 . doi: 10 .1007/

[18] F. Bonanno , G. Capizzi, S. Coco , C. Napoli , A . Lau- 978 -3- 642 -29347- 4 _3. dani, G. L. Sciuto, Optimal thicknesses determi- [29] S. Russo , C. Napoli , A comprehensive solution for nation in a multilayer structure to improve the psychological treatment and therapeutic path planspp eficiency for photovoltaic devices by an hy- ning based on knowledge base and expertise sharbrid fem - cascade neural network based approach, ing , volume 2472 , 2019 , pp. 41 - 47 . 2014 , pp. 355 - 362 . doi: 10 .1109/SPEEDAM. 2014 . [30] N. N. Dat , V. Ponzi , S. Russo , F. Vincelli , Supporting 6872103. impaired people with a following robotic assistant

[19] G. Capizzi , F. Bonanno , C. Napoli , A wavelet by means of end-to-end visual target navigation based prediction of wind and solar energy for long- and reinforcement learning approaches, volume term simulation of integrated generation systems , 3118 , 2021 , pp. 51 - 63 . 2010 , pp. 586 - 592 . doi: 10 .1109/SPEEDAM. 2010 . [31] V. Marcotrigiano , G. D. Stingi , S. Fregnan , P. Mag5542259. arelli, P. Pasquale, S. Russo , G. B. Orsi , M. T. Mon-

[20] F. Bonanno , G. Capizzi, C. Napoli , Some remarks tagna, C. Napoli, C. Napoli , An integrated control on the application of rnn and prnn for the charge- plan in primary schools: Results of a field investidischarge simulation of advanced lithium-ions bat- gation on nutritional and hygienic features in the tery energy storage , 2012 , pp. 941 - 945 . doi:10. apulia region (southern italy) , Nutrients 13 ( 2021 ). 1109/SPEEDAM. 2012 . 6264500 . doi: 10 .3390/nu13093006.

[21] F. Bonanno , G. Capizzi, G. L. Sciuto , C. Napoli, [32] V. Ponzi , S. Russo , V. Bianco , C. Napoli , A. WaG. Pappalardo , E. Tramontana , A novel cloud- jda, Psychoeducative social robots for an healthier distributed toolbox for optimal energy dispatch lifestyle using artificial intelligence: a case-study, management from renewables in igss by using wrnn volume 3118 , 2021 , pp. 26 - 33 . predictors and gpu parallel solutions , 2014 , pp. 1077 [33] G. De Magistris , R. Caprari , G. Castro, S. Russo , - 1084 . doi: 10 .1109/SPEEDAM. 2014 .6872127. L. Iocchi , D. Nardi , C. Napoli , Vision-based holis-

[22] C. Napoli , G. Pappalardo , E. Tramontana , A hy- tic scene understanding for context-aware humanbrid neuro-wavelet predictor for qos control and robot interaction 13196 LNAI ( 2022 ) 310 - 325 . stability 8249 LNAI ( 2013 ) 527 - 538 . doi: 10 .1007/ doi:10.1007/978-3- 031 -08421-8_ 21 .

[34] N. Brandizzi , S. Russo , G. Galati, C. Napoli , Address- 2018. arXiv: 1709 .10082. ing vehicle sharing through behavioral analysis: A [47] D. V. Lu , D. Hershberger , W. D. Smart , Layered solution to user clustering using recency-frequency- costmaps for context-sensitive navigation, in: 2014 monetary and vehicle relocation based on neigh - IEEE/RSJ International Conference on Intelligent borhood splits, Information (Switzerland) 13 ( 2022 ). Robots and Systems , 2014 , pp. 709 - 715 . doi:10. doi: 10 .3390/info13110511. 1109/IROS. 2014 . 6942636 .

[35] S. Pepe , S. Tedeschi , N. Brandizzi , S. Russo , L. Ioc- [48] ROS contributors, nav2: Navigation2 stack for ros2, chi, C. Napoli, Human attention assessment us- ROS2 Ecosystem , 2024 . URL: https://navigation.ros. ing a machine learning approach with gan-based org/# , accessed: 2024 -01- 12 . data augmentation technique trained using a cus- [49] ROS contributors, Ros (robot operating system), tom dataset , OBM Neurobiology 6 ( 2022 ). doi:10. 2024 . URL: https://www.ros.org, accessed: 2024 -01- 21926/obm.neurobiol. 2204139 . 12.

[36] A. Alfarano , G. De Magistris, L. Mongelli , S. Russo, [50] ROS contributors, Collision monitor node, ROS2 J. Starczewski , C. Napoli , A novel convmixer trans- Ecosystem , 2024 . URL: https://https://navigation. former based architecture for violent behavior de- ros .org/configuration/packages/collision_ tection 14126 LNAI ( 2023 ) 3 - 16 . doi: 10 .1007/ monitor/configuring-collision -monitor-node. 978-3-031-42508-0_1 . html, accessed: 2024 -01-12.

[37] C. Ciancarelli , G. De Magistris, S. Cognetta, [51] J. van den Berg , M. Lin , D. Manocha , Reciprocal D. Appetito , C. Napoli , D. Nardi , A gan ap- velocity obstacles for real-time multi-agent naviproach for anomaly detection in spacecraft teleme- gation , in: 2008 IEEE International Conference tries 531 LNNS ( 2023 ) 393 - 402 . doi: 10 .1007/ on Robotics and Automation, 2008 , pp. 1928 - 1935 . 978 -3- 031 -18050-7_ 38 . doi: 10 .1109/ROBOT. 2008 . 4543489 .

[38] I. E. Tibermacine , A. Tibermacine , W. Guettala, [52] J. van den Berg , S. Guy, M. Lin , D. Manocha , RecipC. Napoli, S. Russo , Enhancing sentiment anal- rocal n-Body Collision Avoidance , volume 70 , 2011 , ysis on seed-iv dataset with vision transformers : pp. 3 - 19 . doi: 10 .1007/978-3- 642 -19457- 3 _ 1 . A comparative study, 2023 , pp. 238 - 246 . doi:10. [53] N. Vinukonda , K. N. Shah , Dynamic obstacle 1145/3638985 .3639024. avoidance for assistive autonomy using orca,

[39] G. D. Magistris , C. Rametta , G. Capizzi, C. Napoli , 2024 . URL: https://nagarjunvinukonda.github.io/ Fpga implementation of a parallel dds for wide-band ifles/DR_Obstacle_Report .pdf, accessed: 2024 -01- applications, volume 3092 , 2021 , pp. 12 - 16 . 12 .

[40] G. De Magistris , M. Romano , J. Starczewski, [54] H. Wen , J. Yang, Research on a costmap that can C. Napoli, A novel dwt-based encoder for human change the turning path of mobile robot, Jourpose estimation , volume 3360 , 2022 , pp. 33 - 40 . nal of Physics: Conference Series 2005 ( 2021 )

[41] M. Everett , Y. F. Chen , J. P. How , Motion plan- 012095. URL: https://dx.doi.org/10.1088/ 1742 -6596/ ning among dynamic, decision-making agents with 2005 /1/012095. doi: 10 .1088/ 1742 - 6596 /2005/ deep reinforcement learning, in: IEEE/RSJ In- 1/012095. ternational Conference on Intelligent Robots and Systems (IROS) , Madrid, Spain, 2018 . URL: https: //arxiv.org/pdf/ 1805 . 01956 .pdf.

[42] M. Everett , Y. F. Chen , J. P. How , Collision avoidance in pedestrian-rich environments with deep reinforcement learning , IEEE Access 9 ( 2021 ) 10357 - 10377 . doi: 10 .1109/ACCESS. 2021 . 3050338 .

[43] R. Falque , C. L. Gentil , F. Sukkar , Dynamic object detection in range data using spatiotemporal normals , 2023 . arXiv: 2310 . 13273 .

[44] V. Kurtz , H. Lin , Toward verifiable real-time obstacle motion prediction for dynamic collision avoidance , 2019 . arXiv: 1811 .01075.

[45] P. Long , W. Liu, J. Pan , Deep-learned collision avoidance policy for distributed multi-agent navigation , 2017 . arXiv: 1609 . 06838 .

[46] P. Long , T. Fan , X. Liao , W. Liu, H. Zhang , J. Pan, Towards optimally decentralized multi-robot collision avoidance via deep reinforcement learning,