atharva-pantheon/pilot-corpus-v0.1

Pilot Corpus v0.1

A frozen, pointer-based miniature of a robot-learning pre-training mixture: 7 tranches spanning teleop, UMI, egocentric dexterous manipulation, ego/exo wild video, internet action video, and sim — with every label type (vision / proprio / action / force / pose / lang) and every domain tag represented, sized to run scratch-vs-init experiments on a single GPU.

What this repo is

Manifests + code, not media. A clip is a virtual (episode, t_start, t_end) pointer into a source dataset — we never cut or transcode video. manifests/*.parquet is the single source of truth; the loader reconstructs frames on demand from the original datasets. To use this corpus you download the underlying datasets yourself (see Sources below) and point the loader at them. No third-party video is redistributed here — that keeps the artifact license-clean (several sources prohibit redistribution).

Layout

src/corpus/         schema, manifest I/O, loader (v2.x/v3.0/image-parquet
                    adapters), quota sampler, manifest builder
manifests/          per-tranche + combined parquet manifests (the SoT)
                    + exo_crossview_index.csv, internet_template_index.csv
probes/             frozen eval sets (episode-level splits) + FROZEN_HASHES.txt
configs/quota.yaml  per-batch tranche mixture (a pilot starting point)
tests/              schema round-trip, v3.0 decode, build, sim/internet
scripts/            gen_stats.py, freeze_probes.py, validate_phase0.py
STATS.md            one-page corpus summary (generated)

Manifest schema

One row per virtual clip: clip_id, tranche, source_repo, source_episode, t_start, t_end, native_fps, resolution, lbl_{vision,proprio,action,force,pose, lang}, domain_tags, drift_score, contact_events, pose_eval_only, split. native_fps is preserved verbatim end-to-end (never resampled) — it is the axis the mask scheduler needs for contact-transient eligibility.

Sources (download these yourself)

Every manifest row points into one of the datasets below — no media is stored here. Clip/hour counts are what the manifests currently reference (see STATS.md); "on-disk" is the raw footprint of the local ingest.

tranche	source (HF repo)	slice ingested	clips / hours	labels	fps	license
egodex	griffinlabs/EgoDex-LeRobot-v3.0	test split, 111 tasks, 257 episodes	15,739 / 12.8h	vision, pose, lang	30	CC-BY-NC-ND
umi	lerobot/umi_cup_in_the_wild	full, 1,447 episodes	45,479 / 37.8h	vision, proprio, action	10	permissive
teleop	lerobot/aloha_mobile_cabinet	full, 85 episodes (effort→force)	1,615 / 1.4h	vision, proprio, action, force	50	permissive
exo	hyf015/EgoExoLearn	155 videos (80 ego + 75 exo), 8 tasks	38,310 / 31.9h	vision	~25–30	see dataset
egowild	a1raman/epic_kitchens_100	36 videos, 7 participants	9,247 / 7.7h	vision	30–60	CC-BY-NC
internet	morpheushoc/something-something-v2	100k clips, all 174 templates	179,184 / 138.8h	vision, lang	12	no redistribution
sim	physical-intelligence/libero	1,692 episodes	16,868 / 13.9h	vision, proprio, action, lang	10	permissive

Details of each ingest (participant/task selection, fps histograms, validation) are recorded in raw/<tranche>/_ingest_summary.json.

Staged raw data — downloaded, NOT yet in the manifests

Two AgiBot datasets have been ingested to raw/ since the manifests were last built. They are not yet referenced by any manifest row (so they don't appear in STATS.md or the loader output) — they await LeRobot conversion + a manifest rebuild. They are documented here so the footprint isn't mistaken for stray data.

dir	source (HF repo)	on-disk	status	notes
raw/teleop/AgiBotWorld-Beta	agibot-world/AgiBotWorld-Beta (gated)	72 GB	access granted, downloaded	6 tasks, 8 cams, AV1 @30fps 640×480, per-frame head depth. Carries real 6-DoF end-effector wrench in 88% of episodes (the dataset README's "not available" caveat is outdated). Ships no LeRobot converter — conversion pending.
raw/sim/AgiBotDigitalWorld	agibot-world/AgiBotDigitalWorld	20.7 GB	downloaded	3 tasks, HEVC @30fps, head+hand depth. Wrench empty but joint effort populated (~[-41, 215] Nm) as a force-adjacent signal. Native tar layout; a convert_to_lerobot.py ships with the repo.

Raw on-disk footprint per tranche (incl. the staged AgiBot dirs above): egodex 15G · egowild 46G · exo 16G · internet 8.5G · sim 54G (libero + AgiBotDigitalWorld) · teleop 74G (aloha + AgiBotWorld-Beta) · umi 3.0G.

Quickstart

uv sync
uv run python -m corpus.build_manifest --all     # (re)build manifests from raw/
uv run python scripts/gen_stats.py               # STATS.md
uv run python scripts/freeze_probes.py           # probe sets + hashes
uv run pytest -q                                  # full suite

from corpus.manifest import read_manifest
from corpus.loader import ClipDataset
ds = ClipDataset(read_manifest("manifests/corpus.parquet"), raw_root="raw")
sample = ds[0]        # ClipSample(frames (T,H,W,3) uint8, native_fps, labels, ...)

Known limitations (pilot v0.1)

• AgiBot access is now granted and the data is downloaded, but not yet in the manifests. raw/teleop/AgiBotWorld-Beta (72 GB, real 6-DoF wrench) and raw/sim/AgiBotDigitalWorld (20.7 GB, joint effort) are staged on disk (see Staged raw data above) but still need LeRobot conversion + a manifest rebuild. Until then the manifest teleop tranche is ALOHA only, and the ≥20h contact-rich / ≥10k contact-events targets are not yet met in the SoT.
• Contact pseudo-labels (Phase 2) are not yet computed, so probe_contact is empty.
• Hours are ~10× below the plan's per-tranche targets for the pure-video tranches (the plan's estimates were optimistic); this is a deliberately small, fully-representative pilot, not the full-scale mixture.
• Latent cache (Phase 4) requires a GPU and is not built here.