DA-Next-5M

DA-Next-5M is a collection of egocentric and robot wrist-view data for depth estimation, camera pose estimation, and 3D reconstruction. The full collection described in the accompanying paper contains approximately 5.5 million frames across 22K sequences from seven sources; these reported totals have not been independently recounted for the hosted release.

Data and access

This repository hosts 686 .tar.gz archives under adt/, colosseum/, rlbench/, robolab/, and robotwin/. Archives preserve source-specific image, depth, camera, and metadata layouts, including PNG, NumPy, and JSON files. Decode depth units and camera conventions using the relevant source format. A unified loader and verified train/evaluation split manifest are not currently provided.

Two additional sources are hosted separately:

• HOI4D processed subset: 2,466 archives used in DA-Next-5M, included in the Croissant download inventory.
• Xperience-10M: upstream source requiring manual access approval. The selected 100 sequences and their processed DA-Next-5M outputs are not yet identified by a published download manifest.

Download individual archives from the Files tab or with huggingface_hub. For example, this downloads a small ADT sample:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="HarrisonPENG/DA-Next-5M",
    repo_type="dataset",
    filename="adt/Apartment_release_golden_skeleton_seq100_10s_sample_M1292.tar.gz",
)

Croissant metadata

croissant.json provides version-pinned download URLs, file sizes, SHA-256 checksums, asset schemas, source provenance, and Responsible AI documentation. Local checks passed for JSON, Croissant 1.1, required RAI fields, and loading an ADT sample; the full collection has not been exhaustively validated.

License and intended use

The collection is released under CC BY-NC 4.0. Respect the applicable upstream dataset and asset terms, including the original HOI4D license and Xperience access conditions. Intended for research on embodied 3D perception; validity for safety-critical deployment has not been established. See the Croissant file for limitations, biases, and privacy considerations.