Towards Efficient Image-goal Navigation: A Self-supervised Transformer-based Reinforcement Learning Approach

Abstract

Image-goal navigation is a crucial yet challenging task that requires an agent to navigate to a goal location specified by an image. Modular methods decompose the problem into distinct subtasks and often involve explicit map construction, which can struggle in complex, unstructured environments. In contrast, end-to-end deep reinforcement learning (DRL)-based methods directly output actions from visual input, with recent improvements focusing on enhancing embedding fusion between the current observation and the goal image. However, both approaches fail to fully leverage the rich temporal relationships present in the agent’s visual-action history. In this paper, we address this limitation by employing a self-supervised transformer to predict masked portions of the agent’s visual-action embeddings. To promote spatio-temporal reasoning, a dual-attention shared transformer is utilized for both masked representation learning and policy generation. Our method demonstrates superior performance and generalization ability compared to 12 existing baselines across the Gibson, MP3D, and HM3D datasets. Code and trained models are available at https://github.com/hujch23/DaMVA.