MINT-SJTU/Evo-DexBench_Assets

Evo-DexBench Assets

The redistributable simulation assets for Evo-DexBench, developed by MINT-SJTU. This repository provides the selected object models, scene resources, textures, and task-specific assets used by the benchmark.

✨ Overview

• Ready-to-use runtime layout for the benchmark's asset downloader.
• Selected scene and object resources from Kenney, Poly Haven, YCB, RoboCasa, and NVIDIA IsaacGymEnvs Factory.
• Task-specific geometry and instruction images, with provenance and collection-level license information.
• No PartNet-Mobility models: users obtain those separately through the official gated repository.

This is a simulation asset collection, not a trajectory dataset. It does not include demonstrations, model checkpoints, or training code. It contains benchmark-selected subsets rather than complete copies of the upstream collections.

📦 Contents

The initial bundle is version 2026-09-30, approximately 1.82 GB in uncompressed files. The runtime directory structure is preserved:

Evo-DexBench_Assets/
├── kenney/                 selected prop and brick assets
├── polyhaven/              selected lamps, books, and textures
├── ycb/                    selected YCB object models
├── task_assets/            task-specific models and instruction images
├── maniskill/
│   └── data/
│       ├── assets/mani_skill2_ycb/
│       └── scene_datasets/robocasa_dataset/
├── licenses/               upstream licenses and attribution records
├── LICENSE                 Apache 2.0 for project-authored assets
├── asset_manifest.yaml     collection inventory and checksums
└── README.md

Collection	Contents	License
Kenney	Selected props and brick geometry	CC0 1.0
Poly Haven	Selected scene models and textures	CC0 1.0
YCB	Selected object meshes, textures, and collision models, including the ManiSkill-preprocessed subset	CC BY 4.0
RoboCasa	Selected kitchen fixtures, scene resources, and textures adapted for ManiSkill/SAPIEN	CC BY 4.0
IsaacGymEnvs Factory	Selected reference geometry within task_assets/b04/factory_assets/	BSD 3-Clause
Project-authored assets	Procedural geometry, wrappers, textures, and instruction images	Apache 2.0; adapted third-party components retain their own licenses

Files include OBJ, PLY, STL, GLB, URDF, MJCF/XML, material definitions, and image textures. Preserve their relative paths: model descriptions and materials reference other files in this tree.

The asset manifest records each collection's source, license, runtime path, task/layout usage, and collection or file checksums. Collection SHA256 values hash the sorted UTF-8 records <file_sha256> <relative_posix_path>\n within each collection (sha256sum-lines-v1). The code repository's packaged manifest additionally describes the separately downloaded PartNet models and generated overlays.

🚀 Quick Start

From an Evo-DexBench code checkout, create one simulator environment:

tools/setup_environment.sh --profile sim --name evodex-sim
conda activate evodex-sim

Use the vr profile instead if you also need Quest 3 teleoperation. It includes the simulation dependencies.

Complete benchmark asset setup

First accept the terms and obtain access to sapien-sim/PartNetMobility. Then run:

hf auth login
dexbench assets download \
  --bundle-repo MINT-SJTU/Evo-DexBench_Assets \
  --asset-dir "$PWD/assets"
source assets/activate_evodex_assets.sh
dexbench assets verify --asset-dir "$PWD/assets"

The command installs this bundle, downloads the required official PartNet archives with your own access, and generates the benchmark overlays locally. Run it from the code checkout, not from this asset repository. Source the activation file in each new simulator shell.

Download only this repository

If you only need the redistributable files:

hf download MINT-SJTU/Evo-DexBench_Assets \
  --repo-type dataset --local-dir /absolute/path/to/evodex-assets

This downloads files only. It does not install the benchmark, download PartNet, generate overlays, or create the activation script. Use the benchmark downloader for a complete runtime setup.

🔒 Distribution Boundary

The following are not distributed here:

• Original or modified PartNet-Mobility models.
• Robot URDFs and meshes embedded in the Evo-DexBench Python package.
• Collected demonstrations, training datasets, checkpoints, or model code.
• Local runtime outputs, download caches, or test reports.

Access to this repository does not grant access to PartNet-Mobility or change its license. Each user must accept the official provider's terms. Generated PartNet overlays stay in the user's local asset directory and are not part of this public bundle.

⚖️ Licenses & Attribution

This collection has mixed licenses. There is no blanket Apache 2.0 license for the asset repository: the benchmark code license does not replace the terms governing these models, textures, and images.

Project-authored assets are licensed under Apache License 2.0, Copyright © 2026 MINT-SJTU; see LICENSE. Adapted assets retain the applicable upstream terms; consult the manifest and the records under licenses/. Retain required copyright, license, and attribution notices when redistributing files. A task directory may contain components under several licenses.

The included RoboCasa resources are a selected, adapted distribution. The ManiSkill/SAPIEN adaptation notice is retained inside its dataset directory. Task wrappers and procedural geometry are benchmark-specific additions; upstream geometry remains attributed to its original authors.

🙏 Acknowledgement

We thank Kenney, Poly Haven, the YCB Object and Model Set authors, RoboCasa, ManiSkill, SAPIEN, and NVIDIA IsaacGymEnvs for the resources and tools used by Evo-DexBench.

📄 Citation

📬 Contact