Datasets:

physics-code-transfer-bench
/

cross-scenario-physics-code-transfer

	{
	"@context": {
	"@language": "en",
	"@vocab": "https://schema.org/",
	"citeAs": "cr:citeAs",
	"column": "cr:column",
	"conformsTo": "dct:conformsTo",
	"cr": "http://mlcommons.org/croissant/",
	"rai": "http://mlcommons.org/croissant/RAI/",
	"data": {"@id": "cr:data", "@type": "@json"},
	"dataType": {"@id": "cr:dataType", "@type": "@vocab"},
	"dct": "http://purl.org/dc/terms/",
	"examples": {"@id": "cr:examples", "@type": "@json"},
	"extract": "cr:extract",
	"field": "cr:field",
	"fileProperty": "cr:fileProperty",
	"fileObject": "cr:fileObject",
	"fileSet": "cr:fileSet",
	"format": "cr:format",
	"includes": "cr:includes",
	"isLiveDataset": "cr:isLiveDataset",
	"jsonPath": "cr:jsonPath",
	"key": "cr:key",
	"md5": "cr:md5",
	"parentField": "cr:parentField",
	"path": "cr:path",
	"recordSet": "cr:recordSet",
	"references": "cr:references",
	"regex": "cr:regex",
	"repeated": "cr:repeated",
	"replace": "cr:replace",
	"sc": "https://schema.org/",
	"separator": "cr:separator",
	"source": "cr:source",
	"subField": "cr:subField",
	"transform": "cr:transform"
	},
	"@type": "sc:Dataset",
	"name": "cross-scenario-physics-code-transfer",
	"conformsTo": "http://mlcommons.org/croissant/1.0",
	"description": "A cross-scenario physics-transfer benchmark for stress-testing whether within-scenario compositionality metrics on frozen video features predict cross-scenario physical-property transfer. Contains four Kubric physics scenarios (collision, ramp, flat-drop, elasticity) totalling 1,800 scenes plus a 75-scene matched-visual low-gravity collision variant; pre-extracted frozen features for eight video and image backbones (V-JEPA 2, V-JEPA 2.1, DINOv2-S/L, CLIP ViT-L/14, MAE, SigLIP, VideoMAE); ground-truth per-object tracks; and N-shot adaptation protocols with explicit task/label mappings. Real-video evaluation uses the public Phys101 dataset; we redistribute only feature tensors extracted from it. The headline empirical finding is that high TopSim, PosDis, and causal-specificity bottleneck codes do not generalise across scenarios in the tested bottleneck protocol family. This Hugging Face repository hosts the load-bearing reviewer-inspectable subset (V-JEPA 2 collision features, all four Kubric scenario label files, full reproduction code); the remaining backbone features, scenario videos, GT tracks, and Phys101 features are prepared for an immediate post-acceptance public release.",
	"url": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer",
	"version": "1.0.0",
	"license": "https://www.apache.org/licenses/LICENSE-2.0",
	"datePublished": "2026-05-05",
	"creator": {
	"@type": "Organization",
	"name": "Anonymous (NeurIPS 2026 E&D Track Submission)"
	},
	"citeAs": "Anonymous Authors. A Benchmark for Cross-Scenario Physics-Code Transfer: Compositionality Metrics on Frozen Video Features. NeurIPS 2026 Evaluations & Datasets Track (under review).",
	"isLiveDataset": false,
	"keywords": [
	"physics representation learning",
	"cross-scenario transfer",
	"compositionality metrics",
	"TopSim",
	"PosDis",
	"video foundation models",
	"V-JEPA 2",
	"frozen features",
	"benchmark",
	"Kubric",
	"Phys101",
	"negative result"
	],
	"rai:dataCollection": "Kubric scenarios were rendered using the public Kubric simulator (Apache 2.0) with the PyBullet physics backend at 240Hz substeps. Scene parameters (mass, restitution, friction, drop-height, initial velocity) are sampled on explicit grids documented in the accompanying paper. The 75-scene low-gravity variant uses the same RNG seed as the standard-gravity collision dataset to match per-scene physics-random variables (sphere color, lighting, initial velocity, position jitter) for matched-visual analysis. Phys101 features are extracted from the publicly available Phys101 dataset (Wu et al., BMVC 2016) using the V-JEPA 2 frozen encoder; we redistribute the feature tensors, not the source video.",
	"rai:dataCollectionType": "Synthetic simulation (Kubric/PyBullet) plus derived features from public real-video dataset (Phys101).",
	"rai:dataCollectionRawData": "Raw rendered video frames at 256x256, 48 frames at 24 fps; ground-truth per-object position and finite-difference velocity from PyBullet.",
	"rai:dataCollectionTimeframe": "Rendered in 2025-2026.",
	"rai:dataAnnotationProtocol": "All labels are derived directly from the simulator's ground-truth physical state (mass, restitution, friction, drop-height, initial velocity) and binned into discrete classes (3-class union-binned restitution; 5-class mass-ratio for multi-property training; per-scenario tertile mass for Phys101). No human annotation involved.",
	"rai:dataAnnotationPlatform": "Programmatic (Python).",
	"rai:dataAnnotationAnalysis": "Class-balance and bin-boundary statistics are reported in Appendix A.7 of the paper.",
	"rai:dataAnnotationPerItemTime": "0",
	"rai:dataAnnotationDemographics": "Not applicable (synthetic data + derived features).",
	"rai:dataAnnotationTools": "Custom Python scripts; see code/ in this supplementary bundle.",
	"rai:dataAnnotatorDemographicsDescription": "No human annotators involved.",
	"rai:dataPreprocessingProtocol": "Frozen-feature pre-extraction: each scene is processed as 4 evenly-spaced frames at 256x256, with one forward pass per ViT-L-scale encoder; spatial features are mean-pooled within each tubelet/patch grid to produce (N, T=4, D) tensors per scenario per backbone. Stratified train/holdout splits use 80/20 per primary label class. N-shot target-side stratified subsampling is documented in Section 3.6 of the paper.",
	"rai:dataUseCases": "Evaluating whether within-scenario compositionality metrics (TopSim, PosDis, causal specificity) predict cross-scenario transfer of learned physics codes; benchmarking new sender architectures (slot-attention, EMA-codebook VQ, product-quantised continuous, transfer-aware objectives) on a frozen-feature substrate; stress-testing the assumption that high within-scenario metric values entail abstract reusable structure; isolating the contribution of visual versus dynamics versus scene/task-structure shift via the matched-visual transfer ladder.",
	"rai:dataLimitations": "Four Kubric scenarios with 1-2 spheres; broader scenario diversity (objects, occlusions, multi-agent) would strengthen generality. The metric-transfer sweep is collision-trained and primarily evaluated on collision -> ramp at N=192 with a partial replication on collision -> flat-drop. The n=24 correlational analysis is underpowered (widest bootstrap 95% CI [-0.58, +0.32]); the load-bearing claim is sufficiency, not non-predictiveness. Tested code families are Gumbel-Softmax, tanh-bounded continuous, and a 4-configuration VQ-VAE probe; slot-attention and EMA-codebook VQ are explicit forward directions, not tested. Phys101 per-scenario tertile binning means class boundaries differ across source and target, contributing a label-boundary shift on top of the feature-distribution shift on real video; this is documented and a global-binning diagnostic is also released.",
	"rai:dataSocialImpact": "The benchmark is intended to help researchers more honestly evaluate physics-representation methods and reduce overclaiming of compositionality from within-scenario metrics. We see no immediate negative societal impacts. The released features are derived from frozen public foundation models on synthetic and openly-licensed real-video sources; no personally identifiable information is present.",
	"rai:dataBiases": "All scenes are simulator-generated and contain no human subjects, places, or demographic information. Phys101 is a publicly released physics dataset with no personally identifiable content; we redistribute only V-JEPA 2 features extracted from it.",
	"rai:personalSensitiveInformation": "None. The dataset contains no personally identifiable information.",
	"rai:dataReleaseMaintenancePlan": "The benchmark is hosted on Hugging Face. Updates (additional backbones, additional scenarios, additional pre-extracted features) will follow semantic versioning; current release is v1.0.0. Issues and pull requests are tracked at the Hugging Face dataset repository.",
	"rai:hasSyntheticData": true,
	"distribution": [
	{
	"@type": "cr:FileObject",
	"@id": "vjepa2-collision-features",
	"name": "vjepa2_collision_pooled.pt",
	"description": "V-JEPA 2 frozen features for the 600 collision scenes, mean-pooled to shape (N=600, T=4, D=1024), float32. Loaded with torch.load.",
	"encodingFormat": "application/octet-stream",
	"contentSize": "29604716 B",
	"sha256": "2bee296d965bf6a7d8e47c196575aff422e78788136957d7bc272ea647fb5fc2",
	"contentUrl": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer/resolve/main/features/vjepa2_collision_pooled.pt"
	},
	{
	"@type": "cr:FileObject",
	"@id": "labels-collision",
	"name": "labels_collision.npz",
	"description": "Ground-truth physics labels for the 600 collision scenes: per-scene mass scalars, restitution scalars, mass quantile bins (3-class and 5-class), and restitution quantile bins (3-class and 5-class). Loaded with numpy.load.",
	"encodingFormat": "application/zip",
	"contentSize": "75889 B",
	"sha256": "aaa5f2940f4d7bb822e84873c2674bd2a853feefad7e4988044b2e9294eac1dc",
	"contentUrl": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer/resolve/main/labels/labels_collision.npz"
	},
	{
	"@type": "cr:FileObject",
	"@id": "labels-ramp",
	"name": "labels_ramp.npz",
	"description": "Ground-truth physics labels for the 300 ramp scenes: restitution and friction scalars and bins.",
	"encodingFormat": "application/zip",
	"contentSize": "39858 B",
	"sha256": "142cf1154948e3698f954326c0f7ef33a0b71ea997fd46a4181297e53341b4c6",
	"contentUrl": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer/resolve/main/labels/labels_ramp.npz"
	},
	{
	"@type": "cr:FileObject",
	"@id": "labels-flat-drop",
	"name": "labels_flat_drop.npz",
	"description": "Ground-truth physics labels for the 300 flat-drop scenes: restitution and friction scalars and bins.",
	"encodingFormat": "application/zip",
	"contentSize": "23854 B",
	"sha256": "f1e1b6ad78d53cd4a9ecfc648e3f30af61727f2f6669aa429aec9ab2987a230b",
	"contentUrl": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer/resolve/main/labels/labels_flat_drop.npz"
	},
	{
	"@type": "cr:FileObject",
	"@id": "labels-elasticity",
	"name": "labels_elasticity.npz",
	"description": "Ground-truth physics labels for the 600 elasticity scenes: restitution and drop-height scalars and bins.",
	"encodingFormat": "application/zip",
	"contentSize": "55580 B",
	"sha256": "3e328c84a275f739ce81cbb35cf3dd4ab0550085a6674e6d65fd0321f8d43086",
	"contentUrl": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer/resolve/main/labels/labels_elasticity.npz"
	},
	{
	"@type": "cr:FileObject",
	"@id": "labels-ramp-3prop",
	"name": "labels_ramp_3prop.npz",
	"description": "3-property labels for ramp (multi-property training): joint mass + restitution + friction labels for the 300 ramp scenes.",
	"encodingFormat": "application/zip",
	"contentSize": "38324 B",
	"sha256": "954eea264e382b56dcca1fa2c8c4501fb3fb8c46d485ecf596b985b662ff3a2f",
	"contentUrl": "https://huggingface.co/datasets/physics-code-transfer-bench/cross-scenario-physics-code-transfer/resolve/main/labels/labels_ramp_3prop.npz"
	}
	]
	}