Path Planning on Aerial Image by Imitation Learning

Learn a cost map on aerial image from demonstrations

January 10, 2017 - 1 minute read -
machine-learning

Apprenticeship learning/imitation learning is interesting since for a given task we can learn from demonstrations of human or animals without access to a reward/cost function. Sometimes apprenticeship learning/imitation learning could be converted to learning reward/cost functions.

One of my imitation learning projects was Path Planning on Aerial Image by Imitation Learning. The idea was first to extract image feature maps, and then learn a cost map, which is a linear combination of the feature maps, that minimizes the cost of demonstrated routes. I did fair amount of image processing and handcrafting features in this project. I also implemented gradient descent for learning the cost map.

aerial image

⬆️ Aerial image of university of pennsylvania

aerial image

⬆️ Learned cost map for walking

aerial image

⬆️ Learned cost map for driving

aerial image

⬆️ Example of a planned walk route

aerial image

⬆️ Example of a planned drive route

Check out detailed project report, and presentation slides on github repo