Learning Obstacle Avoidance

The project aims to demonstrate a successful application of Deep Reinforcement Learning to train a quadrotor to learn to avoid obstacles using visual cues. Double Deep Q Networks are employed to train a drone agent in obstacle course simulations to learn how to roll when different kinds of obstacles are encountered. Though the resulting model is suboptimal, it does learn to avoid different types of obstacles provided they are convex and not congested beyond to maneuverability of the quadrotor.

Detailed Report

Project Page