End-to-End Efficient Indoor Navigation with Optical Flow

There has been a recent interest in employing reinforcement learning for training end-to-end goal-driven robot navigation policies. However, implementing reinforcement learning in end-to-end navigation may result in inefficient policies that exhibit redundant turning actions when attempting to avoid obstacles. This work proposes a two branches network to learn efficient policies with less turning action when robots cross the obstacles. We first employ supervised learning to train a robot action classification network with optical flow. We then combine this classifier with an RGBD optical encoder to develop an action-decision network. Ultimately, we evaluate our approach in a visually realistic simulation environment. The results show that our method can reduce unnecessary steering actions and improve efficiency while ensuring navigation capabilities. We further show that our approach can reduce energy consumption during navigation and extend the robot's work time. Experiment results in the iGibson® simulator over hand-made paths reveal that our method can reduce 13.1% of the action number in the training set and 12.9% in the testing set compared with the baseline approaches. It also can reduce 8.3% energy consumption in the training set and 9.6% in the testing set and only has a 4.2% and 8.1% difference compared with the human path.