Visual Language Navigation (VLN) powered navigation robots have the potential to guide blind people by understanding and executing route instructions provided by sighted passersby. This capability allows robots to operate in environments that are often unknown a priori. Existing VLN models are insufficient for the scenario of navigation guidance for blind people, as they need to understand routes described from human memory, which frequently contain stutters, errors, and omission of details as opposed to those obtained by thinking out loud, such as in the Room-to-Room dataset. However, currently, there is no benchmark that simulates instructions that were obtained from human memory in environments where blind people navigate. To this end, we present our benchmark, Memory-Maze, which simulates the scenario of seeking route instructions for guiding blind people. Our benchmark contains a maze-like structured virtual environment and novel route instruction data from human memory. To collect natural language instructions, we conducted two studies from sighted passersby onsite and annotators online. Our analysis demonstrates that instructions data collected onsite were more lengthy and contained more varied wording. Alongside our benchmark, we propose a VLN model better equipped to handle the scenario. Our proposed VLN model uses Large Language Models (LLM) to parse instructions and generate Python codes for robot control. We further show that the existing state-of-the-art model performed suboptimally on our benchmark. In contrast, our proposed method outperformed the state-of-the-art model by a fair margin. We found that future research should exercise caution when considering VLN technology for practical applications, as real-world scenarios have different characteristics than ones collected in traditional settings.

The increasing demand for intelligent systems capable of interpreting and reasoning about visual content requires the development of Large Multi-Modal Models (LMMs) that are not only accurate but also have explicit reasoning capabilities. This paper presents a novel approach to imbue an LMM with the ability to conduct explicit reasoning based on visual content and textual instructions. We introduce a system that can ask a question to acquire necessary knowledge, thereby enhancing the robustness and explicability of the reasoning process. Our method comprises the development of a novel dataset generated by a Large Language Model (LLM), designed to promote chain-of-thought reasoning combined with a question-asking mechanism. We designed an LMM, which has high capabilities on region awareness to address the intricate requirements of image-text alignment. The model undergoes a three-stage training phase, starting with large-scale image-text alignment using a large-scale datasets, followed by instruction tuning, and fine-tuning with a focus on chain-of-thought reasoning. The results demonstrate a stride toward a more robust, accurate, and interpretable LMM, capable of reasoning explicitly and seeking information proactively when confronted with ambiguous visual input.