aras62/task67_hybrid_v4_assets

Task 67 hybrid v4: frozen demonstration pool (assets_pool200n)

This dataset is a static runtime input of task67-v4.0.0 ("hybrid v1"), a policy for BEHAVIOR-1K Task 67 boxing_food_after_dinner (BEHAVIOR Challenge 2026). The policy runs a frozen scripted cabinet stage (frozen_training_v0_86) and then the Task 67 VLA (checkpoint 83346, model repo aras62/vla_67_ft).

These files belong to the frozen cabinet stage, and the policy server does not start without them. At startup the frozen policy reads manifest.json, checks the sha256 of all 200 demonstrations and their 200 wrist references, and stops on any missing or changed file. At run time it picks one demonstration by head-camera matching, then replays and corrects that demonstration's actions. A smaller or changed pool would therefore change the policy. The VLA half needs only its checkpoint.

Files

The dataset holds 412 files in the folder assets_pool200n/: 652,943,580 bytes, about 653 MB. The repository also holds .gitattributes and this card.

Files	Count	Contents
manifest.json	1	Schema task67_pool_v1. Lists the 200 demonstrations (file, sha256, frame count, phase segmentation, wrist file and its sha256) and the sha256 of every other file below.
demo_e<id>.npz	200	One training demonstration each: per-frame states (61 values) and actions (23 values, the R1Pro action space), both float32, for 4,881 to 10,263 frames. Also 32 × 32 head-camera RGB thumbnails at a subset of frames (images, frames).
wrist_demo_e<id>.npz	200	Two 224 × 224 wrist-camera RGB references per demonstration, with their frame indices.
start_views.npz, register_views.npz, resume_views.npz, cabinet_retry_views.npz	4	Views for choosing a demonstration at the start (head RGB and proprioception preselection, then head-depth registration) and for resuming or retrying the cabinet stage.
door_library.json, door_library_views.npz	2	Cabinet door library and its views.
pickup_library.json, pickup_wrist.json	2	Container pickup library and its wrist references.
fridge_anchors.json	1	Fridge anchors.
start_world.json	1	Robot start poses, taken offline from the training instance definitions.
kitchen_surface_map.npz	1	Static kitchen surface map built from training instances.

All of it comes from the Task 67 training demonstrations and training instances. It contains nothing from the public test instances (301–320) or the hidden ones, and the policy never selects by episode ID.

Which release uses it

task67-v4.0.0 pins this dataset at revision 0d5694e5872fcd7e4e729fedad8e388da6a3bf4b, folder assets_pool200n. The release is in the repository markli1hoshipu/Embodied-Agents (private; access is needed), branch task67-hybrid-v4, folder pipelines/behavior1k/task67_boxing_food/releases/v4.0.0. Its README has the results and the setup guide.

Download

Use the release's own script. It downloads the pinned revision, checks every file's sha256, keeps files that already verify, and downloads again any file that fails its hash. From the release folder, with the release's policy environment (see portable/SETUP.md):

.venv-policy/bin/python portable/setup_models.py assets                  # -> assets/assets_pool200n/
.venv-policy/bin/python portable/setup_models.py assets --verify-only    # check only, download nothing

To use another location, pass --assets <folder>/assets_pool200n or set B1K_TASK67_ASSETS. The launchers read the same setting.

To download without the release (no token is needed):

from huggingface_hub import snapshot_download
snapshot_download("aras62/task67_hybrid_v4_assets", repo_type="dataset",
                  revision="0d5694e5872fcd7e4e729fedad8e388da6a3bf4b", local_dir="assets")

Checksums

• Full list: the sha256 of all 412 files is in the release at portable/ASSETS_SHA256SUMS, one <sha256> <file> line per file. To check a copy, run this inside assets_pool200n/: sha256sum -c <release>/portable/ASSETS_SHA256SUMS.
• In the dataset: manifest.json also records the sha256 of every other file, and the policy checks them at startup.
• Against the evaluated runs: these hashes match the input manifests that the evaluated task67-v4.0.0 runs recorded.

License

These files are derived from the BEHAVIOR-1K Challenge training demonstrations for Task 67. The start poses and the surface map come from the challenge's training instance definitions. They are distributed under the same terms as that source data, and using them means accepting the BEHAVIOR-1K dataset terms that apply to it. No other license is granted. The intended use is running and reproducing task67-v4.0.0.