2018-NAVREN-RL: Learning to fly in real environment via end-to-end deep reinforcement learning using monocular images-Gatech best paper award

[jiegenghua]

The main contributions of the paper are as follows: (1)Demonstration of end-to-end Deep RL for collision avoidance using monocular images only and without the use of any other sensing modality. (2)Overcoming the issues associated with the implementation of RL in real environments by designing a suitable reward function that takes into account both the safety and sensor constraints. (3)Using expert data and knowledge-based data aggregation to improve the RL convergence in real time. The dataset is collected by mounting three cameras on a hiker's head facing forward, left and right. A camera frame is taken and is pre-processed before it is fed to the RL system. This pre-processing uses handcrafted algorithms to extract lower dimensional features from the camera image.