These meshes and videos are derived from the Breaking Bad dataset (Sellán et al., 2022), whose everyday subset is built from PartNet / ShapeNet models and inherits the ShapeNet terms of use (https://shapenet.org/terms): non-commercial research and educational use only. Requests are reviewed manually.

Breaking Bad STRIVE

Physics videos of everyday objects breaking and being rebuilt, for STRIVE. Objects from the Breaking Bad everyday subset are placed in texverse-cleverer-strive's emergent scenes, simulated in MuJoCo and rendered in Blender. Some break on impact; others are rebuilt piece by piece by a visible robot tool. Every clip has exact per-instance masks and interaction labels: collisions, contacts, fractures, attaches and holds. This repository also holds what the scenes are made from: the collision proxies and the attach plans.

Layout

collision-proxies/                 the CoACD collision-proxy release everyday-vc-coacd-v1 (its README.md inside)
  assets/<Category>/<object_id>/   complete.obj, complete.proxy.obj, patterns/<pattern>/piece_<i>{.obj,.proxy.obj}
  manifest.jsonl  objects.jsonl  candidates.jsonl.gz  failures.jsonl  provenance.json  checksums.sha256
collision-proxies.zip              the same folder as one ZIP, for downloading
assembly_plans/v5/                 how a scripted tool rebuilds each fracture pattern (its README.md inside)
assembly_plans/v5.zip              the same folder as one ZIP, for downloading
data/breaking-bad-strive-v1/       the dataset's scenes, one ZIP per scene (as they are generated)
  train/<id>.zip  validation/<id>.zip  test/<id>.zip
  <role>/_runs/<fingerprint>/      each run's config.json, inputs.zip (code, inputs, settings) and manifest
data/pilots/<name>/                pilot runs, same layout, kept for comparison

Download a folder as its ZIP. The Hub rate-limits requests, and a folder of thousands of files needs one request per file. Each ZIP holds its folder's files under their repository paths, so unzipping it where you want the repository's root gives you the folder. The folder's checksums.sha256 then checks every file. For example:

hf download strive-vid/breaking-bad-strive collision-proxies.zip assembly_plans/v5.zip --repo-type dataset --local-dir .

The proxy release was at the repository root until 2026-10-11; its files are unchanged under collision-proxies/.

`collision-proxies/`: the collision proxies (`everyday-vc-coacd-v1`)

492 objects with 983 fracture patterns of 2–5 pieces: 3,888 meshes, each normalized (metres, Z-up) and paired with a CoACD compound convex proxy for MuJoCo.

File	Holds
`assets/`	Every complete and piece mesh, and its proxy (one `o convex_NNN` object per hull). Pieces are registered to their complete object.
`manifest.jsonl`	One row per (object, pattern): each piece's mesh and proxy paths with SHA-256, volume, centre of mass, inertia at unit density, repair details, and the adjacency of pieces sharing a fracture wall.
`objects.jsonl`	One row per selected object: status, the complete mesh and its proxy, the normalization transform, the normalized extents, and the selection attempts.
`candidates.jsonl.gz`	Every pattern of the source subset with its piece count: the selection's input.
`failures.jsonl`	Every failed processing attempt (stage and error).
`provenance.json`	The source checksum, seed, selection and CoACD settings, package versions and code revision.
`checksums.sha256`	The SHA-256 of every file of the release, paths relative to `collision-proxies/`.

collision-proxies/README.md has the conventions (frame, proxies, mass properties), the selection and the processing.

`assembly_plans/v5/`: the attach plans

9,830 plans: every pattern for each of 10 tool keys (a parallel-jaw gripper and a suction cup at 0.85×, 1.0× and 1.15×, a Franka Panda hand at 1.0–1.45×). Each plan holds every valid insertion order, the removal paths with tool holds, stable base poses, and the hand-offs that staging needs. 874 of 983 patterns can be rebuilt by some tool. A dataset run reads this frozen set and records its checksums.

`data/`: the scenes

Each ZIP holds one folder, <id>/, with a complete raw dump (schema breaking-bad-strive.raw/v2):

Path	Content
`metadata.json`	The scene: split role, IDs, duration, frame count, image size, counterfactual flags
`frames/rgb.mp4`	Lossless RGB (libx264rgb, CRF 0), every frame, with 1-FPS container timestamps; the render rate is 24 fps
`video.mp4`	An H.264 preview at 24 fps
`segmentations/NNN.npy`	uint16 masks per frame: 0 = background and floor, `i + 1` = instance `i`
`instances/NNN/`	Per instance: source, caption (`category`), poses, velocities, visibility and existence per frame
`camera/`	The camera per frame
`events/events.json`, `events/annotations.json`	Raw contacts, and the labels: collisions, contacts, fractures, attaches, holds, identities
`summary.json`, `background.txt`	texverse's scene summary; the floor material
`breaking_bad/`	The inputs to relabel the scene (scene timeline, physics record, render record)
`_SUCCESS`	Written after the dump validates

Clips: 2–6 s at 24 fps, 512 × 512, rendered with Blender Cycles at 64 samples per pixel. Pilots may say otherwise; v10, for example, is not denoised.
Scene contents: 2–3 resting objects, an initial mover, and objects thrown or slid in. One or two special objects either break at their first visible impact, or are rebuilt by a robot tool and then lifted and tilted. Half of each run's scenes rebuild an object. Some special objects are counterfactual: they do not break, or a piece is let go before it is seated.
Splits are by shape: each Breaking Bad object is in one split by a hash (80/10/10).
Scene IDs count from 0 in every split: test/00000.zip and train/00000.zip are different scenes, drawn from independent seeds (the split is part of every seed).
Captions are noun phrases, such as "wine glass", "wine glass piece" or "wine glass sub-assembly".
SceneSample rows for STRIVE are made from the dumps at training time, by the generator repository's adapter (breaking-bad convert).
Runs: each role's _runs/<fingerprint>/ records the run that made its scenes.
Review sites: the pilots can be browsed at https://justachetan.github.io/breaking-bad-strive/.

Folder	Scenes
`data/pilots/pilot-v10-20261011/`	50: train 40, validation 5, test 5 (2026-10-11)
`data/breaking-bad-strive-v1/`	the dataset (planned 10,000 / 1,000 / 1,000)

Terms of use

This dataset is derived from third-party assets and inherits their terms. Each part is as permissive as its sources allow.

Source	Used for	Terms
Breaking Bad `everyday` (meshes from ShapeNet via PartNet)	breaking and rebuilt objects, their pieces, some distractors, the proxies and plans	the ShapeNet terms of use: non-commercial research and education
texverse-cleverer-strive's approved meshes (mostly Sketchfab CC BY, some CC0)	distractors	each asset's own licence; each scene's `instances/*/instance.json` records its source, licence and artist
Poly Haven and ambientCG materials	floors	CC0 1.0
MuJoCo Menagerie Franka Panda hand	a robot gripper	Apache-2.0

Access is gated with manual approval. The data is provided as is, without warranty. Annotations come from simulation and may contain errors.

Citation

Cite Breaking Bad, PartNet and ShapeNet, and CoACD for the proxies.

@inproceedings{sellan2022breaking,
  title     = {Breaking Bad: A Dataset for Geometric Fracture and Reassembly},
  author    = {Sell{\'a}n, Silvia and Chen, Yun-Chun and Wu, Ziyi and Garg, Animesh and Jacobson, Alec},
  booktitle = {NeurIPS Datasets and Benchmarks Track},
  year      = {2022}
}
@inproceedings{mo2019partnet,
  title     = {PartNet: A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding},
  author    = {Mo, Kaichun and Zhu, Shilin and Chang, Angel X. and Yi, Li and Tripathi, Subarna and Guibas, Leonidas J. and Su, Hao},
  booktitle = {CVPR},
  year      = {2019}
}
@techreport{chang2015shapenet,
  title       = {ShapeNet: An Information-Rich 3D Model Repository},
  author      = {Chang, Angel X. and Funkhouser, Thomas and Guibas, Leonidas and Hanrahan, Pat and Huang, Qixing and Li, Zimo and Savarese, Silvio and Savva, Manolis and Song, Shuran and Su, Hao and Xiao, Jianxiong and Yi, Li and Yu, Fisher},
  institution = {Stanford University, Princeton University, Toyota Technological Institute at Chicago},
  number      = {arXiv:1512.03012},
  year        = {2015}
}
@article{wei2022coacd,
  title   = {Approximate Convex Decomposition for 3D Meshes with Collision-Aware Concavity and Tree Search},
  author  = {Wei, Xinyue and Liu, Minghua and Ling, Zhan and Su, Hao},
  journal = {ACM Transactions on Graphics},
  volume  = {41},
  number  = {4},
  year    = {2022}
}

Downloads last month: 112

Total file size:

32.8 GB

Paper for strive-vid/breaking-bad-strive

ShapeNet: An Information-Rich 3D Model Repository

Paper • 1512.03012 • Published Dec 9, 2015 • 2

Access requires accepting the ShapeNet terms of use