A Spacetime-Aware Synthetic Dataset for Multimodal AI: A Blender-based Framework for Controlled 4D Learning

Title: Spacetime-Aware Multidimensional Datasets for General-Purpose 4D Perception Models

Description:

This paper introduces a conceptual framework and speculative architecture for building 4D datasets that integrate spatial and temporal dimensions into unified representations, enabling general-purpose AI models capable of dynamic perception across time. The work draws inspiration from recent advances in neural radiance fields (NeRF), 3D generative models (e.g., DreamFusion, Nerfies), and spatiotemporal machine learning paradigms (e.g., BEVFormer, Spacetime Neural Networks). It proposes a novel approach to dataset design based on Minkowski spacetime tensors, embedding both video and multimodal spatial inputs into a continuous 4D representation.

We critically address key limitations of prior drafts, including the lack of formal mathematical grounding, vague terminology, and absence of reproducible computational frameworks. This revised version includes references to related state-of-the-art research, outlines a tensorial architecture for encoding spacetime dynamics, and presents a modular structure for constructing 4D benchmarks. Though not implemented as a full pipeline, this work aims to initiate discourse on a next-generation data infrastructure for perception models operating beyond static 3D or sequential 2D paradigms.