HumanCLAW: Can Vision-Language Models Act Through a Body?
Paper • 2607.27180 • Published • 76
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This supplement contains derivative, instance-specific meshes for HSSD-Hab scenes. By requesting access you confirm that you have authorized access to HSSD-Hab, that you comply with the applicable HSSD terms, and that you will use these files for research within those terms.
Log in or Sign Up to review the conditions and access this dataset content.
HumanClawBench HSSD val41 supplement
🌐 Project Page | 📄 Paper (arXiv:2607.27180) | 💻 Code | 🏋️ Motion weights
HumanCLAW evaluates vision-language models as full-body agents in 1,218 find–navigate–interact episodes across 41 HSSD indoor scenes. This gated dataset ships only the small mesh supplement those scenes need. It is not a copy of HSSD and cannot be used on its own.
Step 1 — download the official HSSD data first
The benchmark scenes are built on the official Habitat-ready HSSD dataset
(hssd-hab, version 0.2.5). Request access there and download it before
using this supplement:
- https://huggingface.co/datasets/hssd/hssd-hab (gated; accept the HSSD terms on that page first)
Your download should contain:
/path/to/hssd-hab/
├── hssd-hab.scene_dataset_config.json
├── objects/
├── stages/
└── semantics/
Step 2 — what this supplement adds
Some scene instances cannot be loaded faithfully from the official meshes alone. This supplement provides 1,693 instance-specific baked GLB meshes that:
- bake per-instance scale or reflection into mesh vertices,
- repair triangle winding so Bullet collision behaves correctly, and
- preserve the benchmark cases that intentionally use the render mesh as the exact collider instead of a coarse proxy.
The underlying object geometry comes from HSSD; access is gated and users must comply with the applicable HSSD terms.
Files
hssd/
├── humanclaw-hssd-val41-supplement-v1.tar.gz
└── humanclaw-hssd-val41-supplement-v1.manifest.json
The archive contains 1,693 content-addressed GLB blobs. Their logical size is 184,310,072 bytes (176 MiB); the compressed archive is 83,683,128 bytes.
archive sha256: fd3422b302fcac6696903d73f3d04b54bf66e0603288f3631ca42dd6b4dc8ab2
The manifest maps every HumanClaw instance filename to its exact blob, size, and SHA-256 digest. All 1,693 baked outputs have distinct content hashes.
Step 3 — combine them (automatic)
After accepting access and authenticating with Hugging Face, HumanClawBench downloads, verifies, and caches this archive automatically, then combines it with your official HSSD download:
hf auth login
humanclaw-bench prepare-hssd --hssd-root /path/to/hssd-hab
The original HSSD tree is never modified. The prepared dataset symlinks the official HSSD files and the verified cached supplement.
Offline setup
Download the archive on a connected machine, transfer it to the evaluation host, and pass it explicitly:
humanclaw-bench prepare-hssd \
--hssd-root /path/to/hssd-hab \
--supplement /path/to/humanclaw-hssd-val41-supplement-v1.tar.gz
Passing an already extracted directory containing blobs/ is also supported.
Citation
@article{siyao2026humanclaw,
title = {HumanCLAW: Can Vision-Language Models Act Through a Body?},
author = {Li, Siyao and Gu, Jiawei and Liu, Shuai and Hu, Kairui and Li, Zekun and
Li, Linjie and Tang, Chengcheng and Wu, Po-Chen and Shugurov, Ivan and
Ma, Lingni and Zollhoefer, Michael and An, Sizhe and Mittal, Abhay and
Zhao, Amy and Krishna, Ranjay and Li, Manling and Liu, Ziwei and Guo, Chuan},
journal = {arXiv preprint arXiv:2607.27180},
year = {2026}
}
- Downloads last month
- 13