Visual navigation based on deep reinforcement learning requires a large amount of interaction with the environment, and due to the reward sparsity, it requires a large amount of training time and computational resources. In this paper, we focus on sample efficiency and navigation performance and propose a framework for visual navigation based on multiple self-supervised auxiliary tasks. Specifically, we present an LSTM-based dynamics model and an attention-based image-reconstruction model as auxiliary tasks. These self-supervised auxiliary tasks enable agents to learn navigation strategies directly from the original high-dimensional images without relying on ResNet features by constructing latent representation learning. Experimental results show that without manually designed features and prior demonstrations, our method significantly improves the training efficiency and outperforms the baseline algorithms on the simulator and real-world image datasets.
Keywords: deep reinforcement learning; representation learning; self-supervised auxiliary tasks; target-driven visual navigation.