The dataset viewer is not available for this split.
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
RoboTwin Random MoSE3 supervision cache
Offline MoSE3 supervision for 306,005 windows from 24,922 randomized RoboTwin demonstrations across 50 tasks. The official EEF Random split contains 25,000 demonstrations. The 78 excluded demonstrations each lack a complete set of published camera latents. The exclusions and the affected views are listed in the manifest. Pinned source recordings and per-camera VAE files · File checksums and per-task counts
Use with each backbone
FastWAM-Joint and LingBot-VA can read the same MoSE3 labels for matching task, episode and source-frame keys. FastWAM-Joint encodes three RGB views as a 384 × 320 mosaic (height × width) with its frozen VAE during training. LingBot-VA uses the three per-camera VAE tensors in the official release. This cache contains the teacher targets for the Random demonstrations. The Clean-only training split for Clean→Random has a separate label cache. Clean split MoSE3 labels
Each window contains nine source-frame observations at offsets 0, 4, …, 32 control steps. For each of the eight future offsets, the camera order is head, left wrist, right wrist. Each view provides up to 2,048 points on a 224 × 224 coordinate grid. Each valid point carries 16 MoSE3 feature values. Use point_valid and horizon_valid to select valid entries.
Read one window
Each of the 50 task directories contains an index.json. Its segments object maps <task>/<episode>/<start_frame> to a filename and zero-based row. Together the directories contain 25,315 Safetensors shards. The shards and indexes occupy 526.47 GB.
import json
from pathlib import Path
from huggingface_hub import hf_hub_download
from safetensors import safe_open
repo = "research-vla/robotwin-random-mose3-cache"
task = "adjust_bottle-aloha-agilex_randomized_500-1000"
index_path = hf_hub_download(repo, f"{task}/index.json", repo_type="dataset")
index = json.loads(Path(index_path).read_text())
row = index["segments"][f"{task}/0/0"]
shard = hf_hub_download(repo, f"{task}/{row['file']}", repo_type="dataset")
with safe_open(shard, framework="pt", device="cpu") as tensors:
features = tensors.get_slice("embedding")[row["row"]] # [8, 3, 2048, 16]
point_valid = tensors.get_slice("point_valid")[row["row"]]
The release checks cover the native frame IDs from all three views, every indexed window key, and each shard's row layout and SHA-256. Both backbones' label readers were compared on completed samples from every task. The cache manifest includes the teacher checkpoint fingerprint and the ordered-window digest.
- Downloads last month
- -