Pengembangan Model Navigasi UAV Berbasis Reinforcement Learning untuk Penghindaran Halangan

Unmanned Aerial Vehicle (UAV) mengalami perkembangan pesat dan banyak dimanfaatkan pada misi beresiko tinggi seperti pencarian dan penyelamatan (Search an Rescue), pemetaan, serta patroli wilayah bencana. Namun, keberadaan halangan statis maupun dinamis seperti bangunan, vegetasi, dan objek bergerak masih menjadi tantangan utama dalam navigasi UAV secara otonom. Penelitian ini bertujuan untuk mengembangkan sistem navigasi UAV berbasis Reinforcement Learning (RL) yang mampu melakukan penghindaran halangan secara adaptif dan real-time. Metode yang digunakan meliputi penerapan dan perbandingan beberapa algoritma RL, yaitu Double Deep Q-Network (DDQN), NoisyNet-DQN, dan Q-Learning, yang diuji dalam lingkungan simulasi tiga dimensi menggunakan AirSim. UAV dilengkapi dengan sensor LiDAR dan kamera Depth sebagai sumber data persepsi lingkungan. Hasil penelitian menunjukkan bahwa pendekatan Reinforcement Learning mampu meningkatkan kemampuan navigasi UAV dalam menghindari halangan secara dinamis. Perbandingan antar algoritma menunjukkan perbedaan karakteristik performa dalam hal stabilitas pembelajaran, efisiensi jalur, dan keberhasilan mencapai target. Temuan ini menegaskan bahwa pemilihan algoritma dan strategi eksplorasi berperan penting dalam pengembangan sistem navigasi UAV otonom berbasis Reinforcement Learning.

============================================================================================

Unmanned Aerial Vehicle (UAV) is developing rapidly and widely used in high-risk missions such as Search and Rescue, mapping, and patrolling disaster areas. However, the existence of static and dynamic obstacles such as buildings, vegetation, and moving objects are still the main challenge in autonomous UAV navigation. This research aimed to develop a Reinforcement Learning (RL)-based UAV navigation system that was able to perform adaptive and real-time obstacle avoidance. The methods used the application and comparison of several RL algorithms, namely Double Deep Q-Network (DDQN), NoisyNet-DQN, and Q- Learning, which were tested in a three-dimensional simulation environment using AirSim. The UAV was equipped with a LiDAR sensor and a Depth camera as a source of environmental perception data. The results showed that the Reinforcement Learning approach that was able to improve the navigation ability of UAV in dynamically avoiding obstacles. The comparisons between algorithms showed differences in performance characteristics in terms of learning stability, path efficiency, and success in achieving targets. These findings confirmed that the selection of algorithms and exploration strategies played an important role in the development of autonomous UAV navigation systems based on Reinforcement Learning.