Scaling Vision-and-Language Navigation With Offline RL

• DOI: 10.48550/arxiv.2403.18454
• Publication Date: 2024-03-27
• Authors: Valay Bundele; Mahesh Bhupati; Biplab Banerjee; Aditya Grover
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

The study of vision-and-language navigation (VLN) has typically relied on expert trajectories, which may not always be available in real-world situations due to the significant effort required collect them. On other hand, existing approaches training VLN agents that go beyond data involve augmentations or online exploration can tedious and risky. In contrast, it is easy access large repositories suboptimal offline trajectories. Inspired by research reinforcement learning (ORL), we introduce a new problem setup VLN-ORL studies using demonstration data. We simple effective reward-conditioned approach account for dataset suboptimality agents, as well benchmarks evaluate progress promote this area. empirically various noise models characterizing among unique challenges instantiate VLN$\circlearrowright$BERT MTVM architectures R2R RxR environments. Our experiments demonstrate proposed leads performance improvements, even complex intricate