We present Swarm Manager, a novel distributed computation framework to augment perception accuracy of computationally-constrained platforms in heterogeneous multi-robot systems. The proposed approach exploits computation offloading and software defined networking allowing robots to offload heavy computation (e.g. lidar odometry, object detection) to other more resourceful peers in the team following the decisions of a globally-aware and dynamically-eligible central orchestrator. We demonstrate in simulation enhancements in the perception accuracy of a computationally-constrained platform belonging to an heterogeneous team of four autonomous robots involved in the exploration of a multi-level, GPS-denied and communication-constrained unfinished power plant, gathering real-world data from the Urban Circuit of the DARPA Subterranean Challenge

In this paper, we introduce MAD-FELLOWS, a distributed job offloading framework utilizing Multi-Player Multi-Armed Bandit algorithms. The primary objective of MAD-FELLOWS is to facilitate job offloading processes within networks of Unmanned Ground Vehicles (UGVs). The framework is designed to meet stringent job latency requirements while simultaneously minimizing energy consumption, thereby extending the UGV mission duration. We validate the effectiveness of MAD-FELLOWS through an extensive numerical evaluation, comparing its performance with various baselines, including a centralized, oracle-based approach. Our results demonstrate that: i) MAD-FELLOWS surpasses the baselines, achieving rapid convergence to the performance of the centralized approach in a fully-distributed manner, aligning with latency and energy efficiency criteria; ii) MAD-FELLOWS enhances UGV mission duration by up to 57%, outperforming the baseline approaches.

Multi-UAV systems are considered for many applications, where the UAVs work together as data collection and/or delivery nodes. Therefore, many recent works propose path planning algorithms that consider connectivity as a constraint or optimization objective. However, in most works topological connectivity is targeted or average network performance is optimized. On the other hand, network simulators are used to analyze higher layer protocols for UAV networks with generally randomized or sweeplike mobility. In this work, we analyze the network performance of several connectivity-optimized multi- UAV path planners. We analyze jointly optimized as well as relay assisted UAV networks. Our results show that topologically connected multi-UAV paths do not necessarily lead to acceptable network performance in terms of packet delivery rates and throughput. Mobile relay assisted scenarios perform as well as static relay scenarios with less than half the number of relay nodes. Jointly optimized schemes perform well with low number of UAVs when the transmission ranges are high or the number of nodes is low, where hub nodes are less likely to occur.

The application of Unmanned Aerial Vehicle (UAV) networks to the coverage problem in rural and low-income areas is actively studied nowadays by the research community. A UAV-based network infrastructure is different from traditional cellular networks relying on Base Stations. By mounting small cells on top of UAVs, coverage can be enhanced in regions where network operators avoid to invest due to the low Return on Investment. In case there is a requirement for an enhanced throughput in a rural area (e.g., users requesting IoT applications with strict QoS requriements), a potential solution is to place several UAVs over the same area, thus maximizing the offered throughput. However, this situation would lead to an increase in the energy consumption. To tackle this problem, this paper proposes an optimal solution to the problem of jointly maximizing the offered throughput in rural scenarios where users request microservice-based IoT applications, while minimizing the energy consumption of the swarm of UAVs. A Mixed Integer Linear Programming (MILP)-based formulation is defined and evaluated over realistic scenarios.
Experimental results demonstrate how the solution is able to perform UAV placement in a way to maximize the offered throughput in the highest number of areas, while minimizing the total number of deployed UAVs.

In the realm of 5G wireless, Tactile Internet based applications are envisioned to be supported with the ultra-reliable low-latency communications. Remote telerobotic applications are thus becoming more feasible in mission-critical and safety-critical operations. However, the variations in the outer-loop control over wireless can potentially cause drift in operations, thus disrupting the mission and sometimes can be fatal. Hence jointly optimizing the communication and control for the telerobotic applications is very important. However, in practical scenarios, the telerobotic applications share the wireless channel with other users, many of which are datarate-intensive. To mitigate this problem, we propose to use 5G network slicing to serve the two kinds of users and jointly optimize their datarates to maximize the performance of the high-datarate users while also minimizing the operational errors in telerobotic applications. We formulate this problem with Lyapunov optimization and conducted simulation experiments with practical parameter values. Our results show the advantage of our solution in terms of isolation among the slices and also satisfying the requirements of different applications in each slice.

This paper considers the issue of localizing a lost Autonomous Underwater Vehicle (AUV) using range and position measurements transmitted by an assisting AUV. Revealing the position of the assisting AUV could compromise its own safety and thus we need a privacy-preserving solution to this problem. Since this is a challenging task in the underwater domain, we propose to adopt Time Lock Puzzles, which keep the data of the assisting AUV encrypted until enough sequential work has been performed by the lost AUV. This ensures the safety of the assisting AUV, which gets enough time to move away from the revealed location. The generated delays are handled by augmenting the states of the lost AUV with the state of the lost AUV at the instant it receives a measurement. The augmented states are estimated with an Extended Kalman Filter. The efficacy of the proposed method is presented through a simulation campaign that shows how an AUV can be localized using delayed measurements by leveraging only single-range measurements.

Swarm robotics is a research field on robotic collab- oration where large number of robots are deployed to complete collective real-world tasks. Progress made in the field now tends to be democratized through swarm deployments such that users with different background can learn, program, conduct research or just play with them. One of the main challenges for swarm operators though is to keep the infrastructure at reasonable complexity, maintenance level, and price, to allow occasional users to interact with many robots smoothly. In this paper, we present Qrkey, an open-source library which provides an engaging, accessible and trustworthy user experience to control and track robots in a swarm. After a detailed presentation of the primitives and the protocol proposed by Qrkey, we discuss the security concerns raised by this solution. Finally, we conclude by giving Qrkey limitations and possible future work.

Successfully training models which address the com- plex nature of real environments in Smart Cities and robotic networks is challenging, due to the vast amount of data needed. When employing Reinforcement Learning (RL) models, it is also impossible to let them explore all the actions in the real world, due to potential damages and inappropriate actions. The inherent trial-and-error nature of RL, especially in real-world applications like traffic management, makes an integration with Digital Twins (DTs) attractive: DTs provide a secure environment for iterative training and testing, ensuring the refinement of the models before testing. However, the use of DTs has some significant limitations, as the difference between the model and reality may cause significant risks. This study focuses on the adaptive and self- learning characteristics of this approach, considering a standard navigation task in a highway environment, and analyzes its advantages and potential pitfalls.

Unmanned aerial vehicles (UAVs), known as drones, have gained significant popularity across various military, civilian, and commercial applications. Given the fact that many UAV operations rely on the Global Positioning System (GPS), they inevitably become susceptible to GPS spoofing attacks. In recent years, AI-enabled detection approaches toward UAV GPS spoofing attacks have increasingly received research attention. Therefore, it is crucial to have a systematical understanding of GPS spoofing attacks and collect a comprehensive and quality data set in the construction of effective AI-enabled detection. This paper aims to collect a large dataset of UAV flights under normal and attack scenarios and design an effective detection approach for stealthy UAV GPS spoofing attacks using onboard sensors and machine learning. 30 different features from 4 onboard UAV sensors are extracted in constructing effective AI models. On top of that, we examined different deep learning and machine learning models by fusing important features from our analysis. Our evaluation results in different flight scenarios demonstrated the effectiveness of our proposed approach, in which a high detection accuracy up to 98.7% and a fast detection time of 0.5 second can be achieved using the XGBoost model.

Effective operation and seamless cooperation of robotic systems are a fundamental component of next-generation technologies and applications. In contexts such as disaster response, swarm operations require coordinated behavior and mobility control to be handled in a distributed manner, with the quality of the agents' actions heavily relying on the communication between them and the underlying network. In this paper, we formulate the problem of dynamic network bridging in a novel Decentralized Partially Observable Markov Decision Process (Dec-POMDP), where a swarm of agents, cooperates to form a link between two distant moving targets. Furthermore, we propose a Multi-Agent Reinforcement Learning (MARL) approach for the problem based on Graph Convolutional Reinforcement Learning (DGN) which naturally applies to the networked, distributed nature of the task. The proposed method is evaluated in a simulated environment and compared to a centralized heuristic baseline showing promising results. Moreover, a further step in the direction of sim-to-real transfer is presented, by additionally evaluating the proposed approach in a near Live Virtual Constructive (LVC) UAV framework.

Periodic patroling by Unmanned Aerial Vehicles (UAVs) is crucial for security and monitoring applications including border surveillance, port security, or environmental monitoring. However, existing approaches often fall short when confronted with the intricate demands of these applications. They tend to overlook the heterogeneity in visit frequency requirements across different areas, compromising overall security effectiveness. We contribute a comprehensive formulation of the patrolling problem alongside new evaluation metrics that accurately assess patrolling strategies' performance. We propose a heuristic called Balanced Clusters Patrolling (BCP) which partitions the points of interest into threshold-based clusters and assigns UAVs to match each cluster's demands while meeting the visit frequency constraints. Our extensive simulations under a variety of scenarios reveal that BCP consistently outperforms existing solutions across all significant performance metrics, enhancing performance by up to 92% in reducing total delay.

The communication system is under a paradigm transformation that shifts from traditional bit-level transmission to semantic-level transmission. This transition lays the foundation for complex autonomous driving, necessitating instantaneous processing of substantial data within the constraints of computing capacity and communication bandwidth. In this paper, we propose a novel Task-oriented Source-Channel Coding (TSCC) framework that jointly optimizes source coding and channel coding in a task-oriented manner. Specifically, to reduce communication overhead and guarantee autonomous driving performance, we leverage an autonomous driving agent to guide source-channel coding based on a modified Conditional Variational Autoencoder (CVAE). We test the proposed framework on a well-known autonomous driving platform with different communication channel conditions. The results show that compared to traditional communication and state-of-the-art deep Joint Source-Channel Coding (JSCC), our proposed framework achieves superior performance by saving 98.36% communication overhead and maintains an 83.24% driving score even at 0 dB Signal-to-Noise Ratios (SNR).

The development of the Internet of Things is essential in agriculture to meet the challenges of the agro-ecological transition. However, forwarding the measurements from a massive number of IoT-based sensors distributed in the fields to the internet can be a challenging task, all the more in case of energy, cost, latency, data rate or connectivity constraints. The conventional technologies, as the connection to a cellular network, a gateway or a nano-satellite at low earth orbit, may not always met these constraints. Another approach consists to use a data collector embedded on an Unmanned Aerial Vehicle (UAV). This approach has numerous advantages, as its flexibility and the possibility to reduce the transmit power of the sensor nodes to communicate. However, the trajectories of the UAV have to be optimized beforehand. This paper addresses this issue by modeling the communication ranges of the sensor nodes by hemispheres and by considering the Close Enough Traveling Salesman Problem (CE-TSP) at different flying heights. To solve this problem, an algorithm based on three successive parts, a graph reduction, a partheno-genetic algorithm and heuristic rules, is proposed. This algorithm is first tested on data sets involving a massive number of communicating sensors with various communication ranges, and next on a real agricultural case study. The results highlight the performances of the method proposed and open the way to future perspectives for data collection of IoT-based sensors by means of UAVs.

Early wildfire detection in remote and forest areas is crucial for minimizing devastation and preserving ecosystems. Autonomous drones offer agile access to remote, challenging terrains, equipped with advanced imaging technology that delivers both high-temporal and detailed spatial resolution, making them valuable assets assets in the early detection and monitoring of wildfires. However, the limited computation and battery resources of Unmanned Aerial Vehicles (UAVs) pose significant challenges in implementing robust and efficient image classification models. Current works in this domain often operate offline, emphasizing the need for solutions that can perform inference in real-time, given the constraints of UAVs. To address these challenges, this paper aims at developing a real-time image classification and fire segmentation model and presents a comprehensive investigation into hardware acceleration using the Jetson Nano P3450 and the implications of TensorRT, NVIDIA's high-performance deep-learning inference library, on fire classification accuracy and speed. The study includes implementations of Quantization Aware Training (QAT), Automatic Mixed Precision (AMP), and post-training mechanisms, comparing them against the latest baselines for fire segmentation and classification. All experiments utilize FLAME dataset- a image dataset collected by low-altitude drones during a prescribed forest fire, focusing on key performance metrics such as latency, Mean Pixel Accuracy (MPA), Mean Intersection over Union (MIOU), Frames Per Second (FPS), batch size, throughput, and memory utilization (Active Memory, Allocator State). This work contributes to the ongoing efforts to enable real-time, on-board wildfire detection capabilities for UAVs, addressing speed and the computational and energy constraints of these crucial monitoring systems. The results show a 13\% increase in classification speed compared to similar models without hardware optimization. Comparatively, loss and accuracy are within 1.225\% of original values.