VABench: Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control

VABench. Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control. By Zhongbo Zhang and six others, from the Dalian University of Technology and Nanyang Technological University. Published on the seventeenth of September. Its code repository is worked out: about one hundred and seventy source files, with a readme and a licence.

Spatial intelligence takes more than describing where objects are. When a scene is only partly observed, a model has to work out what evidence is missing, go and get it, interpret it in a common spatial frame, and then act on it. Asking a model about positions in a single image tests only the first sliver of that. The title names three ingredients: visual demonstrations to learn a procedure from, active perception to gather missing evidence, and metric control to act on it.

VA-Bench evaluates the complete loop: observe, reason, act, revise. General-purpose multimodal models learn procedural context from demonstrations that are only RGB images. They choose camera viewpoints actively, issue metric Cartesian commands, and revise those commands from execution feedback. Results report terminal success, nine behavioral diagnostics at the level of trajectories, and progress on subtasks.

The setup withholds shortcuts. Models receive no privileged object poses, no oracle trajectories, and no learned action heads, and a fixed controller that does not depend on the model executes only the targets the model specifies. Whatever succeeds or fails, it is the model's own spatial judgement being tested, not a helper that already knows where things are. VA-Bench contains fourteen base task families, eleven for a single arm and three for two arms, plus seven held-out variants of geometry or layout, and a long-horizon track that composes five objects.

Twelve primary model conditions are each evaluated in three independent runs, over the same twenty physically verified seeds for every base task.

The best-performing model scores one hundred percent on target localization and seventy-eight point nine percent on spatial relations, in the annotated run. Yet its macro-average task success across three runs is only fifty-three point nine three percent, plus or minus three point one seven. Knowing where things are is far from being able to act on it.

Active camera control significantly improves task success over passive multi-view observation. In one matched comparison, success rises from twenty-seven point eight six percent to fifty-seven point five. Being allowed to move the camera roughly doubles success in that comparison. Held-out geometry, on the other hand, can reduce task success by more than thirty percentage points. And no model completes a strict long-horizon episode, despite substantial partial progress.

The benchmark separates what a model can say about space from what it can do in it. Its sharpest results sit in that gap: near-perfect localization alongside roughly half task success, and a clear gain when the model is allowed to go and look for itself.