[HUGGINGFACE]score: 0.36
SpatialAct Benchmark Tests VLM Agents on 3D Spatial Reasoning and Action
May 28, 2026
SpatialAct is a simulator-grounded benchmark probing whether VLMs can translate spatial understanding into actions and refine them through multi-turn feedback in 3D scenes, covering both single-step error correction and multi-turn interactive refinement tasks.
paper
HOW THIS AFFECTS YOU
●
researcherSpatialAct fills a gap between passive spatial perception benchmarks and embodied action tasks, letting you measure whether spatial reasoning actually drives correct behavior in 3D environments.