benchflow/frontierphysics-material-agnostic-visual-inverse-dynamics

FrontierPhysics: Material-Agnostic Visual Inverse Dynamics

Pinned copies of the upstream files used by material-agnostic-visual-inverse-dynamics in FrontierPhysics, hosted by BenchFlow. The task's image build and its verifier download these files from this repository instead of UCLA Box and GitHub. The three PAC-NeRF files are byte-identical to the originals on UCLA Box. The MASIV source is a git archive of commit f05ff177. Its files are identical to that commit on GitHub, but the zip itself was built by BenchFlow and is not an upstream release file.

Contents

File	Bytes	SHA-256	Upstream source
pacnerf/data.zip	2,945,025,303	f6d39f5736350645e543a593e6bca2941b56fbf0d9f694c7dc266cc6e7df8dfe	PAC-NeRF dataset folder on UCLA Box, file f_1624530949617
pacnerf/simulation_data.zip	5,453,120,054	46bd6e24731ae89f6c9ef97db84c2bda04e98d72926edf082d8c1be3dc5cbe0a	Same folder, file f_1624526625558
pacnerf/pytorch_resnet101.pth	237,643,033	006dd158a7acedb75685d688474b00796bba3d6907dfbed2d48b7a081743c77f	Same folder, file f_1624514946129 (BackgroundMattingV2 ResNet-101 weights)
masiv/MASIV-f05ff177b33198023c0980669b6df416767af688.zip	6,380,169	e119ac98354e409b12f85f5ba0e5d1b8e6cd8c4ac4db6104b8fb57487d95dd38	yizhou42/MASIV at commit f05ff177b33198023c0980669b6df416767af688, made with git archive --format=zip -9 --prefix=MASIV-<commit>/

MANIFEST.sha256 lists the same digests in sha256sum -c format.

simulation_data.zip contains the ground-truth particle states. FrontierPhysics uses them only in its verifier. The task's agent sandbox blocks this repository.

Oracle artifacts

oracle/ holds oracle-only material. The task's agent sandbox blocks this repository.

oracle/pretrained/ holds the trained model for each of the task's five scenes. The task's default Oracle loads it: it downloads the package for its assigned scene from a pinned revision of this repository, checks its size and SHA-256, renders the saved simulated trajectory again and recomputes the metrics, without training. Each package is a byte-for-byte copy of the file of the same name in pr241/ of the task author's yizhouz/frontierphysics-oracle-artifacts at revision 14728ead9fe68215bbf7fd3f1f977c8979c153a4. oracle/pretrained/README.md lists them, and oracle/pretrained/LICENSES.md holds their licence notices.

The rest of oracle/ is earlier material: replay packages of retained submissions from the author's revision 9ab8539baad4f84f4a418c60d21ead7462950ab8 and from BenchFlow's H100 runs, with their records. The pretrained Oracle does not use it.

Path	Content
oracle/pretrained/	Trained models for all five scenes, used by the task's Oracle; see oracle/pretrained/README.md
oracle/scene_elastic_0.zip	elastic/0: recovered submission; passed the task verifier in the author's check
oracle/scene_newtonian_0.zip	newtonian/0: recovered submission; passed the task verifier in the author's check
oracle/scene_non_newtonian_0.zip	non_newtonian/0: recovered submission; passed the task verifier in the author's check
oracle/recovered_physical_weights.zip	Recovered constitutive and velocity networks for all five scenes
oracle/recovery_results.json	The author's verifier record for the recovery
oracle/MASIV_training_pipeline_main.zip	Training and evaluation source of the task oracle
oracle/upstream/	The author's README, integration notes, licence notes, manifest and agent denylist
oracle/benchflow-h100/scene_sand_batch_0.zip	sand_batch/0: BenchFlow's run of the task's training Oracle on one H100, with the verifier's output for it; see oracle/benchflow-h100/README.md

oracle/MANIFEST.sha256 and oracle/provenance.json list every file copied from the author's revision 9ab8539baad4f84f4a418c60d21ead7462950ab8 and its upstream path; oracle/pretrained/MANIFEST.sha256 and packages.json cover the pretrained packages; oracle/benchflow-h100/MANIFEST.sha256 lists BenchFlow's package. Source archives keep their original licences, including the GRAPHDECO notices for the Gaussian Splatting code.

Download

Pin a full commit revision. The FrontierPhysics task files record the revisions they use: one for the data files, and one for oracle/pretrained/ in oracle/pretrained_manifest.json. The data files have not changed since commit 985abf3c0734b19d653e45a0429d79aac1eead48. This card and MANIFEST.sha256 were added after that commit.

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="benchflow/frontierphysics-material-agnostic-visual-inverse-dynamics",
    repo_type="dataset",
    revision="985abf3c0734b19d653e45a0429d79aac1eead48",
    filename="pacnerf/data.zip",
)

Attribution and terms

• PAC-NeRF data (data.zip, simulation_data.zip): Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang and Chuang Gan, PAC-NeRF: Physics Augmented Continuum Neural Radiance Fields for Geometry-Agnostic System Identification, ICLR 2023. The PAC-NeRF code is MIT-licensed. The authors' Box folder states no licence for the data, and BenchFlow asserts none. The files are copied unchanged, only so that the benchmark can be reproduced. If you are a rights holder and want them removed, open a discussion on this repository.
• Matting checkpoint (pytorch_resnet101.pth): weights from BackgroundMattingV2 (MIT), copied from the PAC-NeRF Box folder.
• MASIV source (masiv/…zip): yizhou42/MASIV, MIT licence (LICENSE.md inside the archive). Zhao et al., MASIV: Toward Material-Agnostic System Identification from Videos, arXiv:2508.01112.
• Pretrained Oracle models (oracle/pretrained/): trained by the task author, Yizhou Zhao, with the MASIV code (MIT) and code derived from Gaussian Splatting, which is under the Inria and MPII GRAPHDECO licence: research and evaluation use only. oracle/pretrained/LICENSES.md has the full notices.

Browse the FrontierPhysics task datasets collection.