README.md

metadata

license: mit
task_categories:
  - robotics
tags:
  - autonomous-driving
  - world-model
  - tanitad
  - ego-driving
  - camera
  - webdataset
size_categories:
  - n<1K
extra_gated_prompt: >-
  TanitDataSet-R is the research tier of TanitDataSet. This public mirror
  carries ONLY the redistributable (ship-tier) subset of R — today that subset
  is 100% comma2k19 (MIT), byte-identical to TanitDataSet-C. Non-commercial
  records, if and when they are ingested, are internal-only and will never
  appear here. Access is request-gated so the maintainer can see who is using it
  and notify consumers of corrections; the gate is an access log, NOT an
  additional license restriction — your rights are the upstream ones.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text

TanitDataSet-R — the research tier (public redistributable mirror)

⚠️ Read this first

This repo is currently byte-identical to TanitDataSet-C.

R contributes ZERO additional records today. Same 90 episodes, same 14 shards, same sha256s, same 72/18 split, same single source (comma2k19, MIT). R = C ∪ NC is the design; today NC = ∅ on the build box, because no non-commercial source is downloaded or adapted yet.

If you already have TanitDataSet-C, downloading this repo gains you nothing. It exists so the research tier has a stable, correctly-scoped home to grow into — not because it currently holds more data.

---

What R is designed to be, and what it is

	designed	as of this release (MEASURED)
definition	R = C ∪ NC — one build, one schema, per-record tier stamp	✅ machinery live
records	C's permissive records + non-commercial corpora	90 — all of them C's
NC records	nuScenes, BDD100K, A2D2, Argoverse 2, KITTI-360, ONCE, DRAMA, Rank2Tell	0 — registered, none downloaded, no adapters
NC value-add over C	maximum scene diversity, HD maps, semantic layers	0 records, 0 GB, 0 %
redistribution	C-subset ✅ · NC-subset ❌ never	this mirror is 100 % C-subset

What this public mirror will and will not ever contain

TanitDataSet-R as an internal view may hold non-commercial records. This public repo carries only the redistributable subset of that view — the export is guarded by a hard scope assertion (allowed_classes={"owned-safe"}, require_commercial_ok=True, forbid_share_alike=True) that refuses egress if a single row falls outside it.

So the honest statement of what this repo is:

Sayood/TanitDataSet-R = the ship-tier subset of R. Today the ship-tier subset is all of R, which is C. When non-commercial records land internally, they do not land here — this mirror stays the ship-tier subset, and the gap between this repo and internal R becomes real rather than zero.

Never re-push R wholesale. The guard enforces this in code; this note records why the guard exists.

---

Contents (identical to TanitDataSet-C)

Episodes	90 (train 72 · val 18)
Sources	comma2k19 (MIT) × 90 — 100 %
License classes present	owned-safe × 90
Non-commercial (nc-research) rows	0
Gated / refuse / share-alike rows	0 / 0 / 0
Shards	14 tar (11 train + 3 val)
Total size	15.93 GB
Frame format	uint8 [T, 9, 256, 256]

Record schema

• frames — uint8 [T, 9, 256, 256] — 3-frame RGB stack (9 = 3×RGB), canonicalized to f_eff ≈ 266 px.
• actions — f32 [T, 2] — (steer, accel).
• poses — f32 [T, 4] — (x, y, yaw, v).
• per-episode metadata incl. sha256 of the frame blob and build_params_hash.

Shard layout, tar member convention, Parquet catalog and the loading snippet are identical to TanitDataSet-C — see its card for the full description and the tarfile / pyarrow examples.

---

Provenance & verification

Verified 2026-07-25 over the actual payload bytes: 14/14 shards, 90 episodes, 90/90 sha256 re-verified against the recorded frame-blob digests, 0 duplicate ids, 0 train/val id overlap, 0 split/path mismatches, catalog ↔ payload bijection exact. Distinct source corpora in the payload: {comma2k19: 90} — nothing else. Machine-readable in LICENSE_VERIFICATION.json.

---

Honest limits

Everything in TanitDataSet-C's honest-limits section applies verbatim, because this is the same data. In particular:

1. R adds nothing to C today — 0 non-commercial records. The headline above is the single most important fact about this repo.
2. 90 episodes is seed-scale, one source, US highway, forward camera only. No urban, intersection, VRU, night or adverse-weather coverage.
3. L2D contributed 0 records — no LeRobot-v3 → 9-channel adapter exists yet (~2–3 engineering days).
4. PhysicalAI-AV is excluded and always will be, from both tiers. It is gated-confidential — NVIDIA-gated, not redistributable, firewalled and recipe-only. The ingestor raises PermissionError on it, so it cannot become a record on any tier, research included. Nothing here is derived from it.
5. Waymo Open / WOD-E2E and Waymax are refused outright — a distinct refuse license class, worse than non-commercial, because their terms follow the trained weights into the model and vehicle operation. They never enter the lake on any tier, research included.
6. Split is episode-level, not route-disjoint; comma2k19 is one commute route re-driven, so train and val can share road segments.
7. No semantic / VLM labels in this release — frames, actions, poses and provenance only.
8. Anonymization is inherited, not re-applied — upstream comma2k19 MIT footage, nothing added.

---

Which repo should I use?

you want	use
commercially-clean data you may ship	TanitDataSet-C
maximum research diversity	C, for now — R has nothing extra yet
to track the research tier as it grows	watch this repo

Citation / attribution

Cite the upstream source; per-source license text ships in NOTICE.

@article{schafer2018commute,
  title  = {A Commute in Data: The comma2k19 Dataset},
  author = {Schafer, Harald and Santana, Eder and Haden, Andrew and Biasini, Riccardo},
  journal= {arXiv preprint arXiv:1812.05752},
  year   = {2018}
}

Built by the TanitAD Phase-A lake pipeline. Provenance travels with the data.