HOVER: Achieving Autonomy

Ammielle WB
4 min readOct 17, 2020


Also known as an unmanned aerial vehicle (UAV), a drone is a self-flying device, without a pilot on board, capable of completing tasks that range from the mundane to the extraordinary.

Traditional means of transportation have come to a halt when it comes to reaching remote areas in short periods of time, delivering products efficiently and thoroughly inspecting stretches of land. UAVs surpass planes, helicopters, cars and buses because they’re uniquely light, inexpensive and convenient.


Once the purpose of the drone has been decided upon, it will have a specific flight time, speed, obstacle sensory range and altitude, as well as require certain sensors and cameras. Hence, the technological features and components will vary. A drone meant to inspect crops will structurally differ from a drone designed to deliver 100kg packages.

To achieve flight and navigation, there are pieces that all UAVs need. They will need a power source, rotors, a frame and propellers to fly. The frame should be made up of light composite materials, as to not weigh the drone down.

Once we understand how drones operate, we can use their current components and features to develop new ones.


In the past, drones were solely associated with military and commercial uses, from performing reconnaissance to delivering Amazon packages. Given their low cost and great flexibility, UAVs have since been deployed in a breadth of other industries.

By providing immense quantities of data, drones can help with:

  • Land development
  • Urban planning
  • Storm forecasting (eg. hurricanes, tornadoes)
  • Disease control
  • Search and rescue missions
  • Wind turbine and solar panel inspection
  • Smart cities and IoT
  • Disease control and epidemiology
  • Security and safety

By traveling quickly and reaching remote areas, drones can deliver:

  • Blood tests and samples
  • Vaccines
  • Medication
  • Organs
  • Life-saving medical supplies
  • Equipment

Drones can greatly reduce delivery time and transportation costs, while improving the accuracy of aerial images for monitoring, research, surveillance and detection.


Machine Learning (ML)

Because drones can capture tons of data with their sensors and cameras, they are of great help to AI systems. This data can then be fed to machine learning models, allowing them to make more accurate predictions. Although most of this abundant data is unstructured, ML models can still process and analyze this information.

Deep Reinforcement Learning (DRL)

DRL is the discipline of ML capable of training machines, enabling drones to become fully-autonomous. Using a reward and penalty approach, deep reinforcement learning models have produced adaptive systems that learn from experiences. The agent (learning algorithm) learns from the data (from sensors and cameras) in its environment (obstacle avoidance), aiming to perform actions that maximize reward.

During flight, there are many unknown obstacles that make planning the drone route very challenging. To avoid collisions, DRL uses data from the UAVs to make decisions from its surroundings.

Deep Learning (DL)

Deep learning allows machines to solve complex tasks with very diverse and unstructured datasets, of which drones supply copious amounts. If the UAV is assigned to counting people or inspecting sites, it should use object detection. That’s why object detection works so well; it’s a computer vision task that locates specific objects with a bounding box and a class label in images and videos. Thanks to object detection, aerial imagery allows for better industrial analysis and inspection of videos and images. The most popular real-time DL algorithms for computer vision are YOLO (You Only Look Once), SSD (Single Shot Multibox Detector) and those of the RCNN (Region-based Convolutional Neural Networks) family.

YOLO algorithm detecting objects in drone

Because UAVs supply abundant data and AI allows drones to become fully autonomous, they have a symbiotic relationship.


Autonomy implies the ability to make one’s own decisions and function without oversight.

Right now, UAVs can’t accomplish their purpose without operator intervention or supervision. A pilot must program their flight path before the takeoff; there’s direct human input. Therefore, they’re not fully autonomous, but merely self-flying or self-piloting.


HOVER is suggesting a novel concept to solve the aforementioned problem of drone dependency on pre-programmed navigation. UAVs would pick up on information about their surroundings, such as other aerial objects (eg. birds, planes), weather, wind direction and obstructive infrastructure (eg. towers, cranes). Once this unstructured data passes through a deep neural network (DNN) model, the drone will decide on its course of maneuver.

Then, the HOVER drones would map the path from their current position to the destination.

Automated ≠ Autonomous

The UAV will either be triggered when an order is placed or a change is detected, or at a regular time on a daily, monthly or weekly basis. Without human input, the drone can accomplish its mission and truly navigate autonomously.

Obstacle avoidance won’t give the UAV autonomy, but individualized launch time and personalized navigation will.


As autonomous vehicles evolve and progress, they are becoming safer, more reliable and convenient. HOVER’s solution is to optimize flight path, to reduce barriers and make UAVs more accessible.