Spatial intelligence unfolds through a perception–action loop: agents act to acquire observations,
and reason about how observations vary as a function of action. Rather than passively processing
what is seen, they actively uncover what is unseen — occlusion, dynamics, containment, and
functionality — beyond the reach of passive sensing.
We take a step beyond prior formulations of spatial intelligence, which often emphasize passive
perception or assume access to oracle observations, by recasting the observer as an actor.
We introduce ESI-Bench, a comprehensive benchmark for embodied spatial intelligence
spanning 10 task categories and 29 subcategories built on OmniGibson, grounded in
Spelke's core knowledge systems. Agents must decide what abilities to deploy — perception,
locomotion, and manipulation — and how to act to answer questions that cannot be resolved
from passive observation alone.
We conduct extensive experiments on state-of-the-art MLLMs and find that
active exploration substantially outperforms passive counterparts, with agents
spontaneously discovering emergent spatial strategies without explicit instruction, while
passive multi-view adds noise rather than signal despite consuming far more images. Most
failures stem not from weak perception but from action blindness, and their coupling
drives cascading failures where bad actions produce bad views which produce worse actions.
While explicit 3D grounding stabilizes reasoning on depth-sensitive tasks, imperfect
reconstruction proves more harmful than 2D baselines by actively distorting spatial relations.
Human studies further reveal that, unlike humans who seek falsifying viewpoints and revise
beliefs under contradiction, models commit prematurely with high confidence regardless of
evidence quality, exposing a metacognitive gap that neither better perception nor more
embodied interaction alone can close.
