Perpustakaan Digital ITB

Payload transportation using Unmanned Aerial Vehicles (UAVs) has been widely used in various applications, such as delivery, construction, and search and rescue operations. While single-UAV systems are limited by the payload’s size and weight, multi-UAV systems can be used to carry larger, heavier, and more uncertain payloads. However, controlling multi-UAV systems to perform specific missions remains a significant challenge; specific coordination and algorithms are required to control these systems. Multi-Agent Reinforcement Learning (MARL) has been a promising approach for enabling more adaptable and practical multi- UAV control and navigation systems. This thesis extends the current decentralized MARL architectures by proposing a cooperative framework for payload transportation. Two independent UAVs are used to carry a payload connected with flexible cables, navigating through a narrow doorway. The width of the doorway is adjusted to be narrower than the length of the rod. The MARL approach aims to encourage cooperation and coordination among agents to carry the payload through the narrow doorway, executing complex asymmetric tilt maneuvers, and reconfiguring the payload’s orientation mid-flight through trial and error. The proposed framework is validated through extensive simulations in Isaac Lab, with results demonstrating complete success without any collisions or failures, within a particular range of initial conditions, indicating viability of the proposed approach for constrained aerial manipulation.