Hugging Face's logo

Datasets:
ChangChrisLiu
/
DisAss

Tasks:
Robotics
Tags:
robotics
vla
world-model
world-action-model
robot-learning
manipulation
License:
Dataset card Files
xet
 Community
DisAss / README.md
ChangChrisLiu's picture
ChangChrisLiu
Add DisAss raw 30 Hz episode ZIP release
5874ca6 verified
preview code
|
Raw
History Blame Contribute Delete
7.01 kB
metadata 
license: cc-by-4.0
pretty_name: DisAss Raw 30 Hz Robot Disassembly Episodes
task_categories:
  - robotics
tags:
  - robotics
  - vla
  - world-model
  - world-action-model
  - robot-learning
  - manipulation
  - disassembly

DisAss: raw 30 Hz robot disassembly episodes

DisAss is a curated release of raw dual-camera robot manipulation episodes for seven computer-component disassembly cases. It contains exactly 50 episodes per case (350 total) selected from a frozen accepted corpus. Every episode is kept in its original SELF-VLA Common V2 representation and packaged as one ZIP.

This is raw acquisition data, not a ready-made 10 Hz, LeRobot, RLDS, or model-specific export. It is intended as source material for end-to-end VLA, world-model, world-action-model, and related robot-learning pipelines whose authors define and validate their own temporal segmentation, actions, splits, and normalization.

Contents

Case Episodes Raw frames Corrected episodes Collection sessions
connector 50 37,550 1 19
cpu 50 90,227 20 13
cpu_fan 50 37,002 1 14
graphic_card 50 56,744 1 17
hdd 50 53,255 0 15
ram 50 34,387 1 17
ssd 50 54,692 1 11
Total 350 363,857 25 —

hdd_limited and ssd_limited are deliberately excluded. The frozen source corpus contains only 29 and 2 accepted episodes for those limited-clearance variants, respectively, so they cannot meet the 50-episode quota and are not silently merged into canonical HDD/SSD.

Exact archive sizes, SHA-256 values, episode/session UUIDs, frame counts, and skill/config hashes are in manifests/episodes.jsonl.

How the data were collected

Each episode follows the SELF-VLA Common V2 acquisition contract:

  1. While the robot is parked, the operator selects and authenticates the component, physical instance, collection intent, skill CSV, and execution configuration for that episode.
  2. L25 arms recording. Joystick motion begins a role-neutral teleop interval.
  3. R34 starts the authenticated procedural skill from the current TCP. CPU can instead use R38 to move to an operator-captured start reference, dwell 0.5 s, and then begin its skill.
  4. Procedural motion is recorded as skill. An evidence-backed intervention may add human correction and linked procedural skill_resume intervals.
  5. The episode is validated and atomically committed before HOME. HOME/camera lifecycle receipts, the physical outcome, and an independent data-quality decision are then attached to metadata.

Saved frames are driven by synchronized camera arrivals: a frame is recorded only when both base and wrist cameras advance, at the slower stream rate (normally 30 Hz). The robot state is aligned to the camera-pair reference time. Teleoperation/control receipts run at 200 Hz and are stored separately; 200 Hz is not the saved image-frame rate.

Repository layout 

README.md
DATA_STRUCTURE.md
manifests/
  episodes.jsonl
  episodes.csv
  summary.json
  source_authorities.json
checksums/
  SHA256SUMS
data/
  connector/*.zip
  cpu/*.zip
  cpu_fan/*.zip
  graphic_card/*.zip
  hdd/*.zip
  ram/*.zip
  ssd/*.zip
tools/
  inspect_episode.py

Each component directory contains 50 ZIPs. Each ZIP contains one unchanged raw episode directory:

episode_YYYYMMDD_HHMMSS_microseconds/
  episode_meta.json
  frame_0000.pkl
  frame_0001.pkl
  ...
  control_trace.jsonl
  causal_events.jsonl

See DATA_STRUCTURE.md for the field-level contract, phase/action semantics, units, and downstream conversion guidance.

Download

Download the small documentation and manifests:

hf download ChangChrisLiu/DisAss \
  --repo-type dataset \
  --exclude "data/**" \
  --local-dir DisAss

Download one case:

hf download ChangChrisLiu/DisAss \
  --repo-type dataset \
  --include "data/ram/*.zip" \
  --include "manifests/**" \
  --include "checksums/**" \
  --include "tools/**" \
  --local-dir DisAss

Download a named episode by using its archive_path from the manifest:

hf download ChangChrisLiu/DisAss \
  data/ram/episode_YYYYMMDD_HHMMSS_microseconds.zip \
  --repo-type dataset \
  --local-dir DisAss

Inspect and extract

Metadata inspection does not load pickle frames:

python DisAss/tools/inspect_episode.py \
  DisAss/data/ram/episode_YYYYMMDD_HHMMSS_microseconds.zip

Explicitly inspect one trusted frame's keys, types, shapes, and dtypes:

python DisAss/tools/inspect_episode.py \
  DisAss/data/ram/episode_YYYYMMDD_HHMMSS_microseconds.zip \
  --frame 0

Safely extract after validating all member paths and types:

python DisAss/tools/inspect_episode.py \
  DisAss/data/ram/episode_YYYYMMDD_HHMMSS_microseconds.zip \
  --extract extracted

Python pickle can execute code. Only unpickle data obtained from this trusted repository, and do so in a controlled environment. The inspection tool loads a pickle only when --frame is explicitly supplied.

Selection policy

Candidates had to be canonical, operator-accepted, fps=30, complete on disk, and strict passes for schema, image/robot stream, 10 Hz derivability, robot state variation, and TCP/joint integrity. The historical field robot_10hz_ok means the 30 Hz raw feedback can safely support a derived 10 Hz representation; it does not mean these source files are 10 Hz.

Within each case, selection is deterministic from seed DisAss-raw-v1-2026-08-28. It preserves a proportional correction stratum (at least one when available) and samples round-robin across collection sessions using SHA-256 rankings. Full rules and source-ledger hashes are in manifests/source_authorities.json.

Important use boundaries

  • No train/validation/test split is assigned. Downstream users should prevent leakage across collection sessions and any other grouping appropriate to their experiment.
  • Segment at episode, phase, control-owner, timing, and repair boundaries before downsampling or constructing next-step actions.
  • commanded_action is a human teleop/correction receipt. It is null on procedural skill and skill_resume frames; null does not mean a zero action.
  • Robot state is observed state. A control receipt is an attempted/issued command. Do not treat the two as interchangeable targets.
  • Outcome, collection intent, correction decisions, and data acceptance are episode-level provenance. They are not automatically valid predictor inputs.
  • This release provides source evidence, not rewards, returns, policy targets, graph labels, future outcomes, or Action-Expert proposals.

Integrity and license

Verify downloaded archives from the repository root:

sha256sum -c checksums/SHA256SUMS

The license metadata already present in this Hugging Face repository is retained: CC BY 4.0.