New Dataset and Training Boost Embodied Agent Dialog Navigation

A new research initiative aims to significantly advance the capabilities of embodied agents in understanding and generating dialog for physical interaction, particularly within indoor navigation scenarios. The existing DialNav framework, which evaluates the complete dialog-execution loop in photorealistic environments, has been hampered by a severe lack of training data, with only 2,000 episodes available. To overcome this data scarcity, the researchers developed an automatic generation pipeline to construct RAINbow, a massive training dataset comprising 238,000 episodes for DialNav. This pipeline efficiently converts existing Vision-and-Language Navigation (VLN) datasets into high-quality, multi-turn dialog formats. Complementing this data expansion, two key methodological advances were introduced: Dual-Strategy Training, a navigation scheme designed to align training with the dynamic dialog-navigation loop, and a specialized localization model that leverages existing VLN knowledge. The combination of the RAINbow dataset and these innovative training strategies has led to substantial performance improvements. The new model significantly outperforms previous baselines, achieving an 89% increase in success rate on 'Val Seen' and a 100% increase on 'Val Unseen' splits, thereby establishing a new state-of-the-art for embodied dialog navigation.