README.md

---
license: mit
task_categories:
  - image-segmentation
tags:
  - robotics
  - simulation
  - deformable-linear-objects
  - isaac-sim
  - synthetic-data
---

# DeformX-Assets (minimal bundle)

Source/render assets for [**DeformX**](https://github.com/DeformX/DeformX), a co-simulation
framework for deformable linear objects (DLOs) that couples a Cosserat rod engine with
NVIDIA Isaac Sim.

These are the **inputs** used to render the
[WireSeg-36k](https://huggingface.co/datasets/DeformX/WireSeg-36K) dataset — scene USDs,
wire CAD assets, rod trajectories, and environment textures. If you want the *rendered
images*, you want WireSeg-36k instead; this repo is for regenerating or extending the data
yourself.

Total size is ~2.1 GB. Everything needed to reproduce the paper's renders is here except
the NVIDIA datacenter scene and the Texturelabs ground textures, neither of which we are
permitted to redistribute (see [What is not here](#what-is-not-here)).

## Usage

```bash
hf download DeformX/DeformX-Assets --repo-type dataset --local-dir ./asset_wireseg36k

cd /path/to/DeformX
export DEFORMX_ASSET_ROOT=/abs/path/to/asset_wireseg36k
python Dataset_generator/cli.py --variant 0 --frame 0 --do_seg --do_depth
```

`DEFORMX_ASSET_ROOT` is the only variable you need — see the DeformX README for the rest.

### Smaller download

Most of the 2.1 GB is HDRIs and datacenter trajectories. For a ~200 MB subset that still
runs the drop/hang generator end to end, restrict the download:

```bash
hf download DeformX/DeformX-Assets --repo-type dataset --local-dir ./asset_wireseg36k \
  --include "usd/*" --include "wires_traj_data/*" --include "ground/*" \
  --include "scripts/*" --include "*.md" --include "*.json" \
  --include "background/boiler_room_4k.hdr" \
  --include "background/machine_shop_01_4k.hdr" \
  --include "background/empty_workshop_4k.hdr" \
  --include "background/small_workshop_4k.hdr"
```

`--include` takes one pattern per flag. Passing several patterns to a single `--include`
makes the CLI treat all but the first as positional filenames and silently skip them.

Both the HDRI and ground-texture samplers randomize over whatever files are present, so a
partial download still renders — you just get less appearance diversity.

## Contents

| Path | Size | What it is |
|---|---|---|
| `usd/rod_drop_multi_{2,4,8}_plane.usdc` | 34 KB | Drop scenes (2/4/8 wires) |
| `usd/rod_hang_flying.usdc` | 9 KB | Hang scene |
| `usd/wire_usdc/wire_usdc/*.usdc` | 21 MB | 72 wire meshes — 12 colors × 3 radii (`r0.001`/`r0.003`/`r0.005`) × 2 finishes (smooth/rough) |
| `wires_traj_data/*.npz` | 8.1 MB | Rod trajectories — `drop_n{2,4,8}` are 100 frames, `hang_n{2,4,8}` are 200 |
| `datacenter/data center camera {1..4}.usdc` | 32 KB | Datacenter camera rigs |
| `datacenter/data center lights.usdc` | 4 KB | Shared datacenter light rig |
| `datacenter/data center cable_*.usdc` | 7.5 MB | Datacenter cable meshes (5 colors) |
| `datacenter/data_grid_clean/dgrid_c{1..4}_ns{04,08,16}.npz` | 844 MB | All 12 datacenter trajectory configs — camera index `cX`, wire count `nsYY` |
| `background/*.hdr` | 1.2 GB | All 47 HDRI environment maps used in the paper |
| `ground/*.jpg` | 38 MB | 18 ground-plane textures, `easy`/`mid`/`hard` difficulty tiers |

`manifest.json` lists every file with its SHA-256 and size.

## What is *not* here

**The NVIDIA datacenter scene.** The full datacenter renders in the paper use
`Datacenter_NVD@10012` (~9.6 GB), an NVIDIA Omniverse content pack governed by the
[NVIDIA Omniverse License Agreement](https://docs.omniverse.nvidia.com/platform/latest/common/NVIDIA_Omniverse_License_Agreement.html).
We cannot redistribute it. Obtain it from NVIDIA and point `--datahall_usd` at
`.../Assets/DigitalTwin/Assets/Datacenter/Facilities/Stages/Data_Hall/DataHall_Full_01.usd`.

The **drop/hang** generator needs nothing from NVIDIA and runs fully from this bundle.

**35 of the 53 ground textures.** The paper randomizes over 53 ground textures. The 18
[Pexels](https://pexels.com) ones are included; the other 35 are from
[Texturelabs](https://texturelabs.org), whose license does not permit redistributing the
raw files. Download those directly from Texturelabs if you want the paper's exact
randomization. The texture sampler randomizes over whatever is present, so the included 18
still render fine.

All 47 HDRIs *are* included (Poly Haven, CC0). `scripts/fetch_full_backgrounds.py` is kept
as a convenience for re-fetching them at a different resolution (`--resolution 1k|2k|8k`).

## Provenance and licensing

| Group | Origin | License |
|---|---|---|
| `usd/`, `wires_traj_data/`, `datacenter/*.usdc`, `datacenter/data_grid_clean/` | Authored by the DeformX authors | MIT, as with the DeformX repo |
| `background/*.hdr` | [Poly Haven](https://polyhaven.com) | CC0 |
| `ground/*pexels*.jpg` | [Pexels](https://pexels.com) | Pexels License |

Original photographer names are preserved in the `ground/` filenames.

## Citation

```bibtex
@inproceedings{deformx,
  title     = {DeformX: A Versatile Co-Simulation Framework for Deformable Linear Objects},
  author    = {Yang, Yi and Fei, Xiang and Wang, Lehong and Li, Chenhao and Dai, Zilin and Kou, Henry and Li, Lu and Choset, Howie},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026}
}
```