Regularized Denoising Masked Visual Pretraining for Robust Embodied PointGoal Navigation

Embodied PointGoal navigation is a fundamental task for embodied agents. Recent works have shown that the performance of the embodied navigation agent degrades significantly in the presence of visual corruption, including Spatter, Speckle Noise, and Defocus Blur, showing the weak robustness of the agent. To improve the robustness of embodied navigation agents to various visual corruptions, we propose a navigation framework called Regularized Denoising Masked AutoEncoders Navigation (RDMAE-Nav). In a nutshell, RDMAE-Nav mainly consists of two modules: a visual module and a policy module. In the visual module, a self-supervised pretraining method, dubbed Regularized Denoising Masked AutoEncoders (RDMAE), is designed to enable the Vision Transformers (ViT)-based visual encoder to learn robust representations. The bidirectional Kullback–Leibler divergence is introduced in RDMAE as the regularization term for a denoising masked modeling task. Specifically, RDMAE mitigates the gap between clean and noisy image representations by minimizing the bidirectional Kullback–Leibler divergence. Then, the visual encoder is pretrained by RDMAE. In contrast to existing works, RDMAE-Nav applies denoising masked visual pretraining for PointGoal navigation to improve robustness to various visual corruptions. Finally, the pretrained visual encoder with frozen weights is applied to extract robust visual representations for policy learning in the RDMAE-Nav. Extensive experiments show that RDMAE-Nav performs competitively compared with state of the arts (SOTAs) on various visual corruptions. In detail, RDMAE-Nav performs the absolute improvement: 28.2% in SR and 23.68% in SPL under Spatter; 2.28% in SR and 6.41% in SPL under Speckle Noise; and 9.46% in SR and 9.55% in SPL under Defocus Blur.