Background

At the core of embodied cognition lies the challenge of understanding and reasoning about space, its representations, dynamics, how it forms from and constrains visual perception, language communication, action, and interaction with the environment. Spatial representations are inherently multimodal, integrating vision, language, and motor control. They must capture both 2D and 3D spatial structures, from local object relationships to global scene geometry. Although we live in a dynamic 3D world, the images and videos we perceive offer only partial, static glimpses of this environment. Reconstructing an accurate, temporally coherent model of the 3D world from such limited observations remains a central challenge in computer vision. Spatial reasoning is equally critical for language understanding, especially in human-robot interaction, where interpreting grounded references to objects and locations is essential for effective communication. For embodied agents, building robust spatial representations is vital to support real-world navigation, manipulation, and dialogue. Despite advances in data-driven pretraining across 2D/3D vision, language, and action modalities, spatial understanding remains an open challenge. For example, (vision-)language models still fall short of human-level spatial reasoning, often struggling to interpret the semantics of spatial expressions, particularly when perspective, memory, or embodiment constraints are involved.

Scope and Goal

This workshop is particularly interested in how spatial representations can be learned from multimodal data, and applied to core tasks in computer vision (CV), natural language processing (NLP), and robotics. Spatial understanding has historically been pursued separately by these communities, each adopting distinct approaches and problem formulations. The goal of this joint workshop is to provide a platform for much-needed cross-disciplinary dialogue, advancing research on spatial understanding and representation by bringing together diverse perspectives. We aim to foster discussion on how spatial representations, whether symbolic, neural, verbal, or geometric, can be learned, evaluated, and deployed across modalities and tasks. A key focus is aligning these approaches with the demands of real-world applications and addressing practical challenges such as the high cost of real-robot experiments, the scarcity of multimodal 3D and embodied data, and the complexities of human-in-the-loop evaluation and non-verbal communication.