Search

Uncertainty of target motion, limited perception ability of onboard cameras, and constrained control have brought new challenges to unmanned aerial vehicle (UAV) dynamic target tracking control. In virtue of the powerful fitting ability and learning ability of the neural network, this paper proposes a new deep reinforcement learning (DRL)-based end-to-end control method for UAV dynamic target tracking. Firstly, a DRL-based framework using onboard camera image is established, which simplifies the traditional modularization paradigm. Secondly, neural network architecture, reward functions, and soft actor-critic (SAC)-based speed command perception algorithm are designed to train the policy network. The output of the policy network is denormalized and directly used as speed control command, which realizes the UAV dynamic target tracking. Finally, the feasibility of the proposed end-to-end control method is demonstrated by numerical simulation. The results show that the proposed DRL-based framework is feasible to simplify the traditional modularization paradigm. The UAV can track the dynamic target with rapidly changing of speed and direction. [ABSTRACT FROM AUTHOR]

To achieve the perception-based autonomous control of UAVs, schemes with onboard sensing and computing are popular in state-of-the-art work, which often consist of several separated modules with respective complicated algorithms. Most methods depend on handcrafted designs and prior models with little capacity for adaptation and generalization. Inspired by the research on deep reinforcement learning, this paper proposes a new end-to-end autonomous control method to simplify the separate modules in the traditional control pipeline into a single neural network. An image-based reinforcement learning framework is established, depending on the design of the network architecture and the reward function. Training is performed with model-free algorithms developed according to the specific mission, and the control policy network can map the input image directly to the continuous actuator control command. A simulation environment for the scenario of UAV landing was built. In addition, the results under different typical cases, including both the small and large initial lateral or heading angle offsets, show that the proposed end-to-end method is feasible for perception-based autonomous control. [ABSTRACT FROM AUTHOR]

In this paper, we investigate the problem of unmanned aerial vehicles (UAVs) autonomous tracking moving target with only an airborne camera sensor. We proposed a novel integrated controller framework for this problem based on multi-neural-network modules (MNNMs). In this framework, two neural networks are designed for target perception and guidance control, respectively. The deep learning method and reinforcement learning method are applied to train the integrated controller. The training result demonstrates that the integrated controller can be trained more quickly and efficiently than the end-to-end controller trained by the deep reinforcement learning method. The flight tests with the integrated controller are implemented in simulated and realistic environments, the results show that the integrated controller trained in simulation can easily be transferred to the realistic environment and achieve the UAV tracking randomly moving target, which has a faster motion velocity. The integrated controller based on the MNNMs structure has a better performance on an autonomous tracking target than the control mode that combines with a perception network and a proportional integral derivative controller. [ABSTRACT FROM AUTHOR]