This study presents a decentralized shared actor–critic framework for cooperative multi-robot coverage in continuous two-dimensional simulation. The method combines permutation-invariant local observations, continuous differential-drive control, and reward shaping based on stepwise Hungarian assignment distances, collision penalties, and time efficiency. Homogeneous teams of four, five, and six agents are evaluated in an obstacle-free environment using five independent training seeds. In the final training window, the full reward configuration achieved full-team success rates of 98.2 ± 2.9% for four agents, 85.1 ± 18.0% for five agents, and 96.3 ± 2.0% for six agents, with mean landmark coverage above 96% in all cases. The lower mean in the five-agent setting was associated with higher seed-level variability dominated by one low-success seed. Reward ablations without assignment shaping or collision penalties remained viable, and seed-level tests did not show a statistically significant final-window advantage of the full reward configuration. The full configuration reached the 80% rolling-success threshold earlier in median terms, with the clearest seed-level support in the four-agent setting. Within-environment comparison showed higher full-team success than MADDPG and MAPPO under the matched training horizon and final-window protocol. Deterministic arena-size transfer from $15\times 15$ to $30\times 30$ showed decreasing full-team success as arena size increased, while partial landmark coverage remained higher than strict full-team completion. The results support the method for small homogeneous teams in the tested obstacle-free simulation, while larger teams, external obstacles, aerial-robot dynamics, formal safety guarantees, and hardware deployment remain future work.

В данной статье рассмотрены алгоритмы распознавания станций подзарядки и оптимизации траекторий полета беспилотных летательных аппаратов (БПЛА) с применением методов Reinforcement Learning (RL). Представлены традиционные подходы компьютерного зрения и глубокого обучения для идентификации зарядных станций, а также алгоритмы планирования маршрутов, включая A*, Dijkstra и Rapidly Exploring Random Tree. Проведен сравнительный анализ методов и показано, что RL обладает высокой адаптивностью в реальных условиях, обеспечивая улучшение автономности и энергоэффективности БПЛА. Исследование демонстрирует, как интеграция RL способствует повышению точности стыковки дронов с зарядными станциями и оптимизации их работы в динамической среде.

Learning-enabled control requires training to produce competent policies consistently when success requires every robot to finish. This study evaluates masked entity-attention multi-agent proximal policy optimization (MAPPO) for simulated landmark coverage by five differential-drive robots. Configurations were compared under a fixed reward, optimizer, stopping rule, and 100 held-out scenarios, using trained seeds as statistical units; the actor observes the robot–landmark scene but not the externally computed assignment. Across ten policies per principal configuration, masked entity attention achieved $94.6\pm 4.7\%$ deterministic full-team success, against $9.3\pm 12.4\%$ and $6.8\pm 5.1\%$ for wider (1.21 times its parameters) and ordinary fixed-index controls, every masked-attention seed exceeding every seed of both fixed-index controls. Near-parameter-matched Deep Sets and unmasked-attention controls reached $73.7\pm 14.0\%$ and $72.1\pm 13.1\%$, overlapping it. The same policies retained $91.4$–$96.3\%$ mean success from three to eight robots without retraining, and masked attention remained separated from both fixed-index controls at every rate in a 45-policy learning-rate design. Held-out scenes come from the training distribution; under shift, success fell to $70.9\%$ in an unseen larger workspace, while landmark, sensing, and actuation perturbations kept it above $93.3\%$. Masked entity-attention MAPPO gives the most consistent full-team completion among the tested designs; broader robustness and physical deployment remain unestablished.

The duration of continuous operation of unmanned aerial vehicles (UAVs) is significantly limited by the resource of their onboard batteries. This article presents a comprehensive system that provides a fully autonomous drone landing and subsequent charging of its battery, designed to eliminate this limitation. The proposed solution includes an intelligent charging station equipped with a protective canopy and a multimodal guidance system, as well as a specially designed quadcopter optimized for close integration with this station. The station's architecture combines a vision system (an annular multi-camera installation with a 360° view and active LED markers), wireless communication channels and a high-performance real-time computing module that offloads the drone's onboard systems by transferring resource-intensive computer vision and control tasks to the station. The system implements a multi-stage landing algorithm: GPS-based global navigation is combined with precise visual guidance using convolutional neural networks and fiducial markers.; Deep reinforcement learning is used to plan the approach trajectory under the canopy; at the final stage, an adaptive PID controller is used for positioning with an accuracy of several centimeters. A magnetic docking mechanism is provided for secure fixation of the drone during charging, and security protocols provide the possibility of emergency interruption of the process if necessary. During the simulation and testing of the prototype, the system successfully performed autonomous landing in various conditions (including at night and in moderate winds), demonstrating high positioning accuracy and stable operation. The results obtained indicate that the developed autonomous landing and charging system can significantly increase the duration of UAV flights, while reducing the need for operator participation.