Nav CoT-110K is a large-scale trajectory inference dataset specially constructed for embodied navigation tasks, which was jointly released by Peking University and Shanghai University of Engineering Science in 2025. The related paper results are "Nav-R1: Reasoning and Navigation in Embodied Scenes", which aims to train and evaluate the model's ability to perform structured reasoning and action planning in complex three-dimensional scenes, and serves as an important foundation for the cold start phase of the Nav-R1 framework.

This dataset contains approximately 110,000 step-by-step thought chain trajectories. Unlike traditional navigation datasets that only provide instructions and target locations, this dataset explicitly incorporates structured reasoning processes corresponding to multimodal observations, thus bridging perception, language, and action.