Autonomous Drone Obstacle Avoidance

Project description

How to make a Drone avoid obstacles autonomously on its path to a predefined destination using AI?

- What are the available resources, and how can we use them?
- What techniques are available for training an autonomous flying model?
- How can we prepare the data in order to improve accuracy?
- What evaluation criteria are appropriate for assessing the performance of the complete system?

Context

MRR Drones is a startup that aims to offer Drones as a Service - like Software as a Service (SaaS). Example use cases are Precision Farming, Automated Inspection, and Security. The Drone navigates through unknown territories at relatively low altitudes during its operation. This means there are obstacles like trees and other unexpected objects on the drone’s path that vary depending on the drone's territory. Therefore, the drone must autonomously avoid obstacles in its path.

The goal of the project is to use AI to make a drone autonomously avoid obstacles on its predetermined path.

An obstacle is any solid object that a drone could collide with on its path and thereby get damaged or its flight impacted. Obstacles do not need to be avoided in the fastest/optimal way but the model needs to avoid them reliably and keep the drone on its general path (no ridiculous avoidance manoeuvres like flying above it when around is shorter and possible).

Since the drone has a pathfinding module that manages the drone’s controls, the obstacle-avoiding model should predict the next waypoint (coordinates: longitude, latitude, altitude) and feed it to the pathfinding module.

Results

CNN/ANN machine learning solution
- Predict 10 next waypoint for the flight.

Reinforcement learning solution
- Trying to autonomy fly the drone and avoid objects in its path.