This repository acts as a comprehensive, curated collection of state-of-the-art research papers on spatial reasoning within Multimodal Vision-Language Models (MVLMs). It organizes and highlights advancements in this critical AI subfield, offering a valuable resource for researchers and practitioners seeking to understand current frontiers and benchmarks.

How It Works

The project functions as a living survey, systematically categorizing and linking to seminal and recent papers across diverse facets of multimodal spatial reasoning. It covers classical 2D and advanced 3D spatial understanding, scene/layout reasoning, visual question answering, and extends to embodied AI tasks like vision-language navigation. The survey also incorporates novel sensor modalities such as audio and ego-centric video, aiming to provide a holistic view of the field.

Quick Start & Requirements

This repository is a curated list of research papers, not a runnable software project. Contribution is via GitHub Pull Requests to add new papers or resources. No installation or specific hardware requirements are applicable for users of the list itself. Links to related surveys, benchmarks, and specific research areas (3D Vision, Embodied AI, General MLLM, Sound/Audio/Egocentric, Spatial Benchmark) are provided within the README.

Highlighted Details

• Comprehensive coverage spanning 2D and 3D spatial reasoning, scene/layout understanding, and VQA.
• Inclusion of emerging areas like embodied AI (navigation, action models) and novel sensor modalities (audio, ego-centric video).
• Highlights frontiers in Multimodal Large Language Models (MLLMs) and introduces open benchmarks for evaluation.

Maintenance & Community

The project is maintained through community contributions via GitHub Pull Requests. No specific community channels (like Discord/Slack) or roadmap are detailed in the README.

Licensing & Compatibility

The project is licensed under the MIT License, which generally permits broad use, modification, and distribution, including for commercial purposes, with attribution.

Limitations & Caveats

As a curated list, the repository's "completeness" is dependent on community contributions. It focuses on research papers and benchmarks, not on providing a unified software framework or pre-trained models for direct implementation. The "state-of-the-art" is a moving target, and the survey's recency depends on ongoing updates.