Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models

  • 2025-05-22 18:59:39
  • Runsen Xu, Weiyao Wang, Hao Tang, Xingyu Chen, Xiaodong Wang, Fu-Jen Chu, Dahua Lin, Matt Feiszli, Kevin J. Liang
  • 0

Abstract

Multi-modal large language models (MLLMs) have rapidly advanced in visualtasks, yet their spatial understanding remains limited to single images,leaving them ill-suited for robotics and other real-world applications thatrequire multi-frame reasoning. In this paper, we propose a framework to equipMLLMs with robust multi-frame spatial understanding by integrating depthperception, visual correspondence, and dynamic perception. Central to ourapproach is the MultiSPA dataset, a novel, large-scale collection of more than27 million samples spanning diverse 3D and 4D scenes. Alongside MultiSPA, weintroduce a comprehensive benchmark that tests a wide spectrum of spatial tasksunder uniform metrics. Our resulting model, Multi-SpatialMLLM, achievessignificant gains over baselines and proprietary systems, demonstratingscalable, generalizable multi-frame reasoning. We further observe multi-taskbenefits and early indications of emergent capabilities in challengingscenarios, and showcase how our model can serve as a multi-frame rewardannotator for robotics.

 

Quick Read (beta)

loading the full paper ...