Drone Delivery System

Inspiration

The future use of drones in healthcare also is very thought provoking. The whole industry would be transformed by the use of this technology to improve safety and care delivery. The rapid delivery of vaccines, medications and supplies right to the source could quash outbreaks of life-threatening communicable diseases. In the future, small indoor drones could deliver medicine to the bedside of a patient from the pharmacy, thus eliminating some human steps. This would lead to more rapid and less error prone administration of medications. Nurses and pharmacists can work more efficiently as supplies can be summoned to the bedside instead of the time consuming task of gathering necessary items. Thus, safe and robust drone delivery systems would lead to a future with more outpatient care and even home-based care that used to be delivered in the hospital.

What it does

We create an environment simulating a city with multiple air traffic zones at various altitudes. We deploy a number of drones into the environment. The drone has to travel from its source to the destination aiming for:

• Drones to take the most optimal path (shortest distance) while simultaneously,
• Minimizing the airspace violations: Any particular zone of the system cannot allow more than a threshold number of drones to fly through it. Thus, our drone delivery system can deliver products to the destination in a safe and reliable way.

How I built it

We build the drone simulator using ML-agents framework in Unity 3D which allows us to define customized game-based environments and train Reinforcement Learning (RL) algorithms to train and test your simulations!

Challenges I ran into

We ran into tons of challenges while doing this project ranging from setting up the environment which required specific compatible versions of Unity and ML-Agents, along with designing the game to be based in the RL framework so that it can easily trained using the ML-Agents framework. It took us a lot of iterations to design how the agents detect when they are in certain zones, and also design a set of environments on which we can test.

Accomplishments that I'm proud of

We successfully created a simulation where we deployed “n=50” drones flying from the source to destination. We compared the random walk algorithm (where the drones randomly fly from the source to destination) with the shortest distance algorithm (where we optimized the path of the drone via RL and showed that the optimized algorithm caused all the drones to not only reach their destination in optimized time but also minimize capacity violations.

What I learned

We learnt how to use the Unity framework to simulate an environment and design RL algorithms with key Deep RL techniques.

What's next for Drone Delivery System Using Reinforcement Learning

New algorithms can be now easily developed for our environment and easily trained making benchmarking easier for drone delivery based environments, to make drone delivery systems even more accurate and safe.