The authors built a physical drone that flew 11,000+ random trajectories - mostly crashing into things. They then trained a neural network with the camera image

The authors built a physical drone that flew 11,000+ random trajectories - mostly crashing into things. They then trained a neural network with the camera images, and got surprisingly decent obstacle avoidance and exploration out of it. Paper here: https://arxiv.org/abs/1704.05588 https://arxiv.org/abs/1704.05588