Deep learning is recently showing outstanding results for solving a wide variety of robotic tasks in the areas of perception, planning, localization, and control. Its excellent capabilities for learning representations from the complex data acquired in real environments make it extremely suitable for many kinds of autonomous robotic applications. In parallel, Unmanned Aerial Vehicles (UAVs) are currently being extensively applied for several types of civilian tasks in applications going from security, surveillance, and disaster rescue to parcel delivery or warehouse management. In this paper, a thorough review has been performed on recent reported uses and applications of deep learning for UAVs, including the most relevant developments as well as their performances and limitations. In addition, a detailed explanation of the main deep learning techniques is provided. We conclude with a description of the main challenges for the application of deep learning for UAV-based solutions.

This paper presents a complete system for automatic recognition and diagnosis of electrical insulator strings, which efficiently combines different deep learning-based components to build a versatile solution to the automation problem of the power line inspection process. To this aim, the proposed system integrates one component responsible for insulator string segmentation and two components in charge of its diagnosis. The insulator string segmentation component consists of a novel Fully Convolutional Network (FCN) architecture, termed Up-Net, which enhances the capabilities of the state-of-the-art U-Net network by introducing new skip connections at certain levels of the architecture. We further propose a second variant of the Up-Net network by training it within a Generative Adversarial Network (GAN) framework. The capabilities of the proposed Up-Net variants are incremented by the application of data augmentation and transfer learning techniques, achieving accurate segmentation of the insulator string elements (i.e., discs and caps). Regarding the insulator string diagnosis, we design a Convolutional Neural Network (CNN) which takes as input the mask generated by the insulator string segmentation component and is capable of identifying the absence of a variable number of discs. The second diagnosis component consists of a novel strategy which integrates a Siamese Convolutional Neural Network (SCNN) designed for modeling the similarity between adjacent discs, allowing the detection of several types of disc defects using the same model. The proposed system has been extensively evaluated in several video sequences from real aerial inspections of high voltage insulators, showing robust insulator recognition and diagnosis capabilities.

Deep learning techniques for motion control have recently been qualitatively improved, since the successful application of Deep Q- Learning to the continuous action domain in Atari-like games. Based on these ideas, Deep Deterministic Policy Gradients (DDPG) algorithm was able to provide impressive results in continuous state and action domains, which are closely linked to most of the robotics-related tasks. In this paper, a vision-based autonomous multirotor landing maneuver on top of a moving platform is presented. The behaviour has been completely learned in simulation without prior human knowledge and by means of deep reinforcement learning techniques. Since the multirotor is controlled in attitude, no high level state estimation is required. The complete behaviour has been trained with continuous action and state spaces, and has provided proper results (landing at a maximum velocity of 2 m/s), Furthermore, it has been validated in a wide variety of conditions, for both simulated and real-flight scenarios, using a low-cost, lightweight and out-of-the-box consumer multirotor.

Recent object detection studies have been focused on video sequences, mostly due to the increasing demand of industrial applications. Although single-image architectures achieve remarkable results in terms of accuracy, they do not take advantage of particular properties of the video sequences and usually require high parallel computational resources, such as desktop GPUs. In this work, an inattentional framework is proposed, where the object context in video frames is dynamically reused in order to reduce the computation overhead. The context features corresponding to keyframes are fused into a synthetic feature map, which is further refined using temporal aggregation with ConvLSTMs. Furthermore, an inattentional policy has been learned to adaptively balance the accuracy and the amount of context reused. The inattentional policy has been learned under the reinforcement learning paradigm, and using our novel reward-conditional training scheme, which allows for policy training over a whole distribution of reward functions and enables the selection of a unique reward function at inference time. Our framework shows outstanding results on platforms with reduced parallelization capabilities, such as CPUs, achieving an average latency reduction up to 2.09x, and obtaining FPS rates similar to their equivalent GPU platform, at the cost of a 1.11x mAP reduction.