Past opportunities

Current systems for landing UAVs on ships tend to be GPS or radar based and do not use passive sensing. This project aims to develop a visual guidance solution for maritime launch and recovery of UAVs which is not reliant on specialist hardware on the ship and is resilient to failed or degraded GPS. Moreover, this project will target small vessels that otherwise would not have a UAV capability owing to the lack of infrastructure that can be installed and the significant ship motion caused by operations in anything but calm seas. This will enable UAV support of operations from small uncrewed vessels as well as larger crewed vessels.

The project will investigate the best way to achieve a visually guided approach and landing on to the moving deck of a maritime vessel. It will compare classical machine vision and deep learning approaches to tracking the ship and deck markings with consideration of fusing other sensor modalities such as LiDAR. Use of ship motion prediction via machine learning will also be explored to decide the best time to conduct launch and recovery and to enhance the smoothness of the landing trajectory.

Current systems for landing UAVs on ships tend to be GPS or radar based and do not use passive sensing. This project aims to develop a visual guidance solution for maritime launch and recovery of UAVs which is not reliant on specialist hardware on the ship and is resilient to failed or degraded GPS. Moreover, this project will target small vessels that otherwise would not have a UAV capability owing to the lack of infrastructure that can be installed and the significant ship motion caused by operations in anything but calm seas. This will enable UAV support of operations from small uncrewed vessels as well as larger crewed vessels.

The project will investigate the best way to achieve a visually guided approach and landing on to the moving deck of a maritime vessel. It will compare classical machine vision and deep learning approaches to tracking the ship and deck markings with consideration of fusing other sensor modalities such as LiDAR. Use of ship motion prediction via machine learning will also be explored to decide the best time to conduct launch and recovery and to enhance the smoothness of the landing trajectory.

Swarms of UAVs are an effective means with which to achieve an objective requiring greater redundancy or greater area coverage than can be provided with a single UAV. However, controlling a large number of UAVs becomes problematic for the operator and places great strain on communication networks. State of the art drone shows are a good example of using large numbers of small drones to great effect but rely on preplanning the trajectories of all the drones before flight to ensure zero collisions and are also reliant on augmented GPS positioning systems to ensure drones fly precise trajectories that do not conflict with one another. This approach does not work in environments having a military context which might be highly dynamic or largely unknown prior to launch and will also not work in GPS degraded environments.

This project aims to solve these limitations by developing neural network architectures capable of outputting control actions which enable a UAV to emulate the sophisticated flocking behaviour of birds. Whilst existing approaches require relative position and velocity of neighbouring UAVs as input, we propose to replace this explicit state information using the raw camera feed from each UAV. The outcomes will contribute to the advancement of communication-free flocking behaviour and pave the way for the adoption of vision-based swarm control in real UAV systems.

Swarms of UAVs are an effective means with which to achieve an objective requiring greater redundancy or greater area coverage than can be provided with a single UAV. However, controlling a large number of UAVs becomes problematic for the operator and places great strain on communication networks. State of the art drone shows are a good example of using large numbers of small drones to great effect but rely on preplanning the trajectories of all the drones before flight to ensure zero collisions and are also reliant on augmented GPS positioning systems to ensure drones fly precise trajectories that do not conflict with one another. This approach does not work in environments having a military context which might be highly dynamic or largely unknown prior to launch and will also not work in GPS degraded environments.

This project aims to solve these limitations by developing neural network architectures capable of outputting control actions which enable a UAV to emulate the sophisticated flocking behaviour of birds. Whilst existing approaches require relative position and velocity of neighbouring UAVs as input, we propose to replace this explicit state information using the raw camera feed from each UAV. The outcomes will contribute to the advancement of communication-free flocking behaviour and pave the way for the adoption of vision-based swarm control in real UAV systems.

This project aims to improve on existing algorithms for SLAM with a view to implementation on Edge AI enabled devices installed on real drones. The desired outcome is a method to intelligently fuse sensors like LiDAR, cameras and inertial systems to provide robust obstacle avoidance capability in cluttered environments without reliable GPS data.

The project will be conducted in both simulation and on real UAV platforms with hardware-in-the loop experiments. Algorithms will be developed for Edge AI enabled devices to achieve UAV payloads that are lightweight, use minimal power whilst meeting flight safety requirements.