luca0621/robocasa-spice-immediate-100

Spice storage: observe storage, then retrieve immediately

This snapshot contains 100/100 verified Expert trajectories (complete). There are no intervening chores. These are training demonstrations, not learned-model successes or memory-benchmark qualification. No fine-tuning has been run.

Known input shortcut: a single threshold on storage-video duration classifies the storage side for 62/100 training episodes. This is a descriptive fit on these same metadata, not a held-out result or learned-policy score. This does not establish that the current view is ambiguous. Use these as manipulation/grounding development data. Do not attribute subsequent fine-tuning success to memory without held-out duration, input-length, and current-view controls. See input-shortcut-audit.json. Raw frames and native timestamps are preserved; no padding, retiming or filtering has hidden this issue.

Each episode includes the full storage observation sequence and subsequent retrieval. Raw three-camera 256x256 RGB, 16-D state, original 20 Hz timestamps, 12-D actions, and an action-loss mask are stored in episodes/<case>/training.hdf5. The MP4 files in reviews/ are human previews; do not train on their overlays.

Training boundary

Let b = first_current_native_step, N = native_actions. Storage observations are [0,b). For every training step t in [b,N), the model may see storage RGB, self-observations through t, the current state/RGB, and the instruction. Only retrieval actions receive loss. Never supply future observations, storage actions, target-side metadata, or scenario/provenance labels as policy input. The final observation at N has no corresponding action. Keep original timestamps.

from hf_dataset import ImmediateEpisode, read_manifest
root = "/path/to/downloaded/dataset"
row = read_manifest(root)[0]
episode = ImmediateEpisode(root + "/" + row["hdf5_path"])
sample = episode.read_sample(row["first_current_native_step"],
                             instruction=row["instruction"], action_horizon=16)
# Use your checkpoint's sampler to supply explicit reference_indices/self_indices.
# Defaults retain raw 20 Hz frames; this reader does not run a VLA processor.

The action order, binary control encoding and state order are documented in schema.json. Normalization has not been applied. The format is native HDF5 with a Hub-readable JSONL episode index, not a LeRobot release or a drop-in SimpleMemVLA trainer. Use the causal reader and the checkpoint's preprocessing when integrating fine-tuning.

Diversity and split

The full plan is 50 paired scenario groups per family, with opposite storage locations. Target objects, distractor placement and initial robot offsets vary within one kitchen layout/style (11/11). Partial snapshots may be imbalanced. All episodes belong to training. Generate held-out groups for evaluation; do not treat a changed motion path of the same scene as an independent test.

Verification and provenance

Only production cases with Expert task success, separate-process native replay error <=1e-8, exact initial/handoff/final three-camera RGB, and MP4 decode/hash checks are included. Failed attempts remain in the generation archive and are not counted. Automated checks do not constitute full manual visual review. See export-validation.json, checksums.json, and per-case provenance/. There are no live pose setters after reset. No model weights are included.

This custom dataset uses RoboCasa scenes/assets; it is not an official RoboCasa benchmark. This staging snapshot does not assign a new blanket asset license.

Download / upload

Download using huggingface_hub.snapshot_download(repo_id, repo_type="dataset"). The JSONL index can also be loaded with datasets.load_dataset. upload_hf.py performs checksum validation and a dry run by default. After both the count and group checks pass, provide --execute --repo-id OWNER/NAME to upload this family to a private dataset repository by default. Use --public only for an explicitly authorized public repository. No upload happens during export.

Publication

This is a public training-data release. The original private upload was blocked by storage quota; public publication was explicitly authorized. Raw RGB, actions and timestamps are unchanged.