Learning Active Re-Localization with Deep Reinforcement Learning

The ability to localize itself relative to the map is critical for a truly autonomous agent to perform various down-stream tasks. Recent works regarding deep learning have made significant progress towards localization problem. But learning-based methods usually work well around training examples due to the nature of supervised learning. This paper investigates a novel method capable of active re-localization to resolve this major limitation. The proposed method incorporates a learning-based localization model, by using a deep neural network to predict the agent's pose and the training examples only cover a small amount areas of the mapped region, and combines it with a policy model, which makes the agent to actively navigate to near training examples in order to rapidly re-localize its pose using the localization model. We cast the active re-localization problem as an episodic reinforcement learning task. The estimated pose is to be used to compute an intermediate reward signal from the environment, thereby the two model can be integrated seamlessly. For evaluate and quantify the effectiveness of the proposed method purposes, we provide a variety of domains for our simulation studies. The experimental results show that the proposed method can accurately re-localize the agent's pose, and outperforms a state-of-the-art active localization method for the average re-localization time.