Li-Yuan Hsu

Computer Science Department at ETH Zürich

Thursday, November 08, 2018, 14:00

Video conference room R 04 007 in Building 106

Supervisors: Stefan Stevšic ́ Prof. Dr. Otmar Hilliges and Prof. Dr. Moritz Diehl

Abstract

We propose two deep recurrent neural network architectures (reinforcement learning and super- vised learning) to solve quadrotor obstacle avoidance and navigation problems. First, training these neural networks only in simulation environment, they are able to directly transfer into real world without any fine-tuning. Both models achieve navigation tasks with success rate over 90%. Second, we show the generalization ability of these models. Training on few simple en- vironments and transferring directly into unseen complex environments, both models perform navigation success rate up to 90%.