| --- |
| license: other |
| license_name: mixed-upstream-terms |
| viewer: false |
| pretty_name: FrontierPhysics — Multi-Object Visual Inverse Dynamics |
| language: |
| - en |
| tags: |
| - frontierphysics |
| - physics |
| - system-identification |
| - mosiv |
| size_categories: |
| - n<1K |
| --- |
| |
| # FrontierPhysics: Multi-Object Visual Inverse Dynamics |
|
|
| Pinned copies of the upstream files used by **`multi-object-visual-inverse-dynamics`** in |
| [FrontierPhysics](https://github.com/benchflow-ai/FrontierPhysics), hosted by |
| [BenchFlow](https://huggingface.co/benchflow). The task's image build and its verifier |
| download these files from this repository instead of the upstream dataset. |
|
|
| This repository holds the ten two-object scenes the task uses: `2objs/{01,02,04,06,23,24,26,45,56,67}.zip`. |
| Each archive is byte-identical to the same path in |
| [Hanibel/MOSIV](https://huggingface.co/datasets/Hanibel/MOSIV/tree/bc1e42d3587420765965f6a27b92b4d9fe8bb987) |
| at revision `bc1e42d3587420765965f6a27b92b4d9fe8bb987`. `MANIFEST.sha256` lists their SHA-256 |
| digests in `sha256sum -c` format, and the task's `verifier/reference_manifest.json` pins each |
| archive's size and digest. |
|
|
| The archives include the hidden evaluation frames and point clouds. FrontierPhysics uses them only |
| in its verifier, and its image build exposes only the observed interval. 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 ten 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 `pr242/` of the task author's [yizhouz/frontierphysics-oracle-artifacts](https://huggingface.co/datasets/yizhouz/frontierphysics-oracle-artifacts/tree/14728ead9fe68215bbf7fd3f1f977c8979c153a4/pr242) 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 ten scenes, used by the task's Oracle; see `oracle/pretrained/README.md` | |
| | `oracle/scene_01.zip` | Scene 01: retained submission, model and trajectory; passed the task verifier in the author's check | |
| | `oracle/scene_24.zip` | Scene 24: retained submission, model and trajectory; passed the task verifier in the author's check | |
| | `oracle/retained_current_results.json` | The author's verifier record for both packages | |
| | `oracle/MOSIV_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_<id>.zip` | Scenes 02, 04, 06, 23, 26, 45, 56 and 67: BenchFlow's retained run of the task's training Oracle (one H100 per scene; unchanged except scene 56's `iter_cnt`) 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 packages. 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 `98974d4d2e76868b674605dfa6c9856806134181`. This card and |
| `MANIFEST.sha256` were added after that commit. |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| |
| path = hf_hub_download( |
| repo_id="benchflow/frontierphysics-multi-object-visual-inverse-dynamics", |
| repo_type="dataset", |
| revision="98974d4d2e76868b674605dfa6c9856806134181", |
| filename="2objs/01.zip", |
| ) |
| ``` |
|
|
| ## Attribution and licence |
|
|
| The MOSIV data (`2objs/`) is released under the MIT licence by its authors. |
|
|
| - **Pretrained Oracle models** (`oracle/pretrained/`): trained by the task author, Yizhou Zhao, with the MOSIV code (MIT) and code derived from [Gaussian Splatting](https://github.com/graphdeco-inria/gaussian-splatting), which is under the Inria and MPII GRAPHDECO licence: research and evaluation use only. `oracle/pretrained/LICENSES.md` has the full notices. |
|
|
| Please cite MOSIV: |
|
|
| ``` |
| @misc{liu2026mosivmultiobjectidentificationvideos, |
| title={MOSIV: Multi-Object System Identification from Videos}, |
| author={Chunjiang Liu and Xiaoyuan Wang and Qingran Lin and Albert Xiao and Haoyu Chen and Shizheng Wen and Hao Zhang and Lu Qi and Ming-Hsuan Yang and Laszlo A. Jeni and Min Xu and Yizhou Zhao}, |
| year={2026}, |
| eprint={2603.06022}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CV}, |
| url={https://arxiv.org/abs/2603.06022}, |
| } |
| ``` |
|
|
| Browse the [FrontierPhysics task datasets collection](https://huggingface.co/collections/benchflow/frontierphysics-task-datasets-6a9e71712c664f5221b85ed1). |
|
|
