cubed-data-v1 / README.md
nomanas25's picture
r3-2026-08-02: BLE sessions republished with relative timestamps (registration derivable from public data; wall clock remains stripped)
8389a67 verified
|
Raw
History Blame Contribute Delete
7.2 kB
metadata
license: cc-by-sa-4.0
pretty_name: Cubed Solve Captures (cubed-data-v1)
tags:
  - rubiks-cube
  - video
  - computer-vision
  - imu
  - sensor-fusion
  - robotics
task_categories:
  - video-classification
  - other
size_categories:
  - n<1K

Cubed Solve Captures (cubed-data-v1)

35 self-recorded captures of a GAN 12 ui Bluetooth smart cube being solved: portrait 1080×1920 H.264 video at ~120 fps with no audio (3.15 GB total), paired with the cube's own BLE move log, phone IMU samples, and a replay-verified scramble record. The corpus exists for camera-only decode research — reconstructing a solve from ordinary phone video — and is the input set for Cubed Core's Benchmark v0. It is a corpus, not a benchmark: no train/validation/test split is assigned here.

Suites

suite captures what it is
cs 25 Complete solves. Full BLE + IMU. Video covers the whole solve on 24 of 25.
gt 6 Research captures (gt17–gt22). Video ends before the solve's final move on five of six. Video + BLE + scramble record; no IMU.
gtD 4 On-camera scramble from a verified solved start (gtD2–gtD5). The scramble itself is filmed. Full BLE + IMU.

All 35 BLE move logs replay bit-exact to solved on two independent cube engines: starting state, then the recorded move sequence, ends on the solved state. The per-move facelets strings embedded in a log are the cube's own periodic state broadcast and lag the move stream, so they are not the replay source of truth — use the scramble/start state plus the move tokens.

Layout and fields

One directory per capture under captures/<capture_id>/.

file captures contents
video.mp4 35 Stream-copied from source: bitstream untouched, audio removed, container metadata stripped.
scramble.json 35 Scramble in standard notation, derived start state, replay-verification and truncation status.
rights_record.json 35 Per-artifact rights binding by byte count and SHA-256.
cube_session.json 35 BLE move log: moves, facelet states, quaternions. moves[].t_ms and orientations[].t_ms are present as integer-millisecond deltas from session start; wall-clock time is still removed.
imu.json 29 Phone IMU orientation samples.
app_cube_session.json 15 App-side BLE log, verified move-identical to cube_session.json.
capture_manifest.json 14 Capture-app manifest.
clip_ble_ground_truth.json 1 Frame-indexed BLE ground truth (gtD1 only).

dataset/manifest.json is the authoritative inventory; SHA256SUMS lists a digest for every published file. Every JSON validates against a schema in the schemas/ directory of the cubed-core repository.

As of this release, cube_session.json adds relative timestamps. moves[].t_ms and orientations[].t_ms are integer-millisecond deltas from the start of the session. This makes frame registration (aligning the move log to the video) computable from public data alone, for every capture that ships a BLE session. Wall-clock date, time of day, and every private ingest handle stay out of the corpus, unchanged from the prior release. The relative timestamps follow the same disclosure the corpus already made for gtD1: clip_ble_ground_truth.json has shipped a frame-indexed BLE record with its own scrub block since the first release, and this update brings the other captures' BLE sessions to the same standard.

No calibration or camera-metadata artifact ships per capture. The corpus was recorded with one cube and one camera setup family, and decode uses the single shared six-colour calibration (calibration_gan12.json) distributed as a cubed-core release asset. That calibration was measured under one lighting condition while the corpus spans three recording cohorts — treat colour evidence away from the gtD1 session accordingly.

Benchmark v0 results

Measured on this corpus; per-row tables, per-capture narratives and hash-bound receipts are under benchmark/.

suite metric result
gt suite reach-LL 5 / 6
gtD1s reach-solved 1 / 1
cs suite reach-LL 4 / 24

gt19 is reported informationally and excluded from the tallies: it was used to train an upstream verifier, so any result on it is contaminated.

Reach-LL asks whether the raw predicted state path reaches last-layer onset with the correct pre-last-layer state, up to whole-cube orientation. These figures come from the evaluation lane, which supplies a ground-truth terminal; the productized camera-only path is a separate lane with its own contract and is not tabulated.

Intended use

This corpus is for camera-only cube-decode research: reconstructing a solve from ordinary phone video, with the cube's own BLE move log available as ground truth. It is also the input set for Cubed Core's Benchmark v0. It is a corpus, not a benchmark — no train/validation/test split is assigned here, and results measured on it are within-solver, within-setup findings.

Decoding requires a CUDA host. The pipeline, the runner and the authoritative configuration live in cubed-core — see docs/tutorials/DECODE.md for the workflow and config/decode-runtime-v1.json for the profiles of record: local_camera_v1 (camera-only, CFG_HASH 1852738634) and canonical_eval_reference (evaluation lane, CFG_HASH 2587091255).

from huggingface_hub import snapshot_download
snapshot_download("cubed-core/cubed-data-v1", repo_type="dataset")

Licenses and attribution

Released under CC-BY-SA-4.0. Attribution: Manas (https://github.com/KingBobJoeIV). Every capture carries a rights_record.json binding each published artifact by byte count and SHA-256, with its own licence and attribution; those per-artifact licences are authoritative. The collection licence in LICENSES/DATASET.md covers selection and arrangement only.

Privacy

All captures were self-recorded by the dataset author solving a GAN 12 ui smart cube filmed with an iPhone, using the project's own capture tooling. No other person appears in or contributed to the recordings. The videos have no audio track and no container metadata: hands and the cube on a desk are the whole frame. The published sidecars carry move tokens, facelet states, quaternions, scrambles and frame indices — no absolute wall-clock times, filesystem paths, device serial numbers, hostnames or account identifiers.

Limitations

  • One solver, one cube, one camera setup family. These captures cannot establish generalization across users, cameras, cubes, grips or lighting.
  • Video ends before the recorded solve's final move on six captures (five of gt17gt22, plus gtD4); each capture's scramble.json records its own truncation status. The BLE log is complete regardless.
  • No per-capture colour calibration ships; decode uses one shared calibration measured under a single lighting condition across a corpus spanning three recording cohorts.
  • The Benchmark v0 figures above come from the evaluation lane, which supplies a ground-truth terminal. They are not camera-only results.