Deep Reinforcement Learning for Goal Directed Visual Navigation

Safe navigation in cluttered environments is a key capability for drone autonomy. This talk presents a Deep Reinforcement Learning based approach for monocular vision based obstacle avoidance. The focus of the work is on training a neural control module that is easy to integrate into conventional control hierarchies and can extend the autonomous capabilities of existing autopilot software stacks. We demonstrate learning neural policies in a visually highly realistic virtual forest environment.

Mate Kisantal, Deep Reinforcement Learning for Goal Directed Visual Navigation