Research map / Vision-language and multimodal / CLIP and zero-shot

AmaraSpatial-10K: A Spatially and Semantically Aligned 3D Dataset for Spatial Computing and Embodied AI

NeurIPS 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Open-vocabulary alignment from VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces AmaraSpatial-10K, a 10K-asset synthetic 3D dataset with metric scale, anchoring, PBR maps, collision hulls and rich descriptions, plus an evaluation suite (SPS, Concept Density, anchor error, CLIP coherence) and downstream retrieval, Holodeck, and physics studies.

Problem

Web-scale 3D asset collections have arbitrary metric scale, incorrect pivots, brittle geometry and incomplete textures, limiting their use in embodied AI, robotics and spatial computing.

Approach

Author spatially aligned (metric scale, canonical front, semantic anchor) and semantically aligned (multi-sentence text, reference image, mesh) synthetic assets, and measure them with a continuous Scale Plausibility Score and an LLM-based concept-density metric rather than relying on post-hoc filtering.

Key results

• CLIP Recall@5 of 0.612 for AmaraSpatial-10K vs 0.181 for Objaverse (3.4x), median rank 267 to 3.
• Mean SPS 0.815 for AmaraSpatial-10K vs 0.412 for matched Objaverse across nine scale-audited categories.
• 99.1% Habitat-Sim stability with V-HACD collision (vs 95.5% for Objaverse) and roughly 20x lower wall time.

Surprising findings

• Objaverse matched seating assets have a mean height of 717.79 m, showing extreme scale outliers in raw community assets.
• Stripping queries to brief descriptions still gives 2.3x Objaverse R@5, showing content rather than render style drives gains.

Limitations

• Assets are procedurally generated rather than scanned, so geometry and material fidelity may be biased relative to photogrammetry.
• Holodeck study is a qualitative case study with N=6 scenes per pack and overlapping standard deviations.
• Metric dimensions are LLM-estimated and CLIP-based scores are only relatively interpretable.

Experimental setup

Datasets: Objaverse HSSD ABO GSO AmaraSpatial-10K

Benchmarks: AmaraSpatial-10K text-to-asset retrieval Holodeck scene composition (6 scenes per pack) Habitat-Sim physics drop test

Metrics: CLIP Recall@1/5/10/25 median rank Scale Plausibility Score coefficient of variation anchor error watertight/manifold rate cross-modal CLIP cosine LLM Concept Density object overlap floor contact

Baselines: Objaverse HSSD ABO GSO Objaverse Matched Default (AllenAI) Holodeck pack

Builds on: Objaverse Holodeck Habitat-Sim CLIP

Closest papers by meaning

• Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation (CVPR 2024 · similarity 0.87)
• EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI (CVPR 2024 · similarity 0.86)
• Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model (ICLR 2025 · similarity 0.86)
• InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.85)
• 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning (CVPR 2025 · similarity 0.85)
• PhysX-3D: Physical-Grounded 3D Asset Generation (NeurIPS 2025 · similarity 0.85)