robospoof / README.md
jeqcho's picture
Publish 900 RoboSpoof paired trial exports and manifest
c2a09ff verified
|
Raw History Blame Contribute Delete
3.82 kB
metadata
language:
  - en
pretty_name: RoboSpoof trial footage
size_categories:
  - n<1K
tags:
  - robotics
  - robot-safety
  - video
configs:
  - config_name: default
    data_files:
      - split: train
        path: manifest.csv

RoboSpoof trial footage

900 robot trials from RoboSpoof, with paired real-camera video and model-observation replays. See the paper.

Downloads

Collection Trials Download
Original 300 original.zip
Blurred 300 blurred.zip
Doctored 300 doctored.zip

Each ZIP includes paired videos, the published model-observation WebP images, per-trial timing files, a manifest, and a README. The combined manifest contains all 900 trials. Its archive and trial-path columns locate each trial within the ZIPs. The URL columns preserve links to the existing Runs release and its sources.

These collections contain separate executions. They are not three matched variants of each recording. Within a trial, the real-camera video and model-observation replay are paired. The headline trial is included in the 900.

What the footage contains

Real-camera videos show the overhead camera at 224×224. Model-view videos replay published observation strips at 672×224, with overhead, left-wrist, and right-wrist cameras in that order. Each camera tile is 224×224. No new spatial crops or overlays are applied to the collection videos.

The available assets are resized and compressed derivatives. They are not full sensor-resolution raw recordings. The ZIPs preserve the exact published WebP bytes for 8,249 observations. MP4s are lossy H.264 re-encodings.

Model-view videos hold each observation until the next one. They are not continuous altered camera recordings and do not include the entire provider request or retained image history. When calls share an action step, the replay shows the last observation at that step. All observations remain in the image files and timing.json.

Playback uses action_step / 10 at 10 frames per second. It omits model inference latency and does not represent elapsed wall-clock time. Timing files preserve the observation steps and playback timestamps. Outcome labels are copied from the published source without new inference.

Headline clips

The matching original headline clip remains on Runs at 224×224. The native doctored headline replay uses 15 native editor outputs before their resize for model input. Those images are 1672×941, with one bottom padding row in the MP4. Its held observations give it less continuous motion than the social demo. The native editor outputs are not the exact resized images sent to the model.

Programmatic download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="robocurve/robospoof",
    repo_type="dataset",
    local_dir="robospoof",
)

The dataset viewer displays manifest rows. The train split name is a loader convention and does not designate a training or evaluation partition.

Provenance and verification

This release mirrors the November 9, 2026 Runs download release. The three ZIPs are unchanged. The combined manifest adds archive locations. Source and video SHA-256 hashes are included in the manifest, and SHA256SUMS covers the files in this release.

No new license grant is asserted by this mirror. Contact Robocurve through the study page for reuse terms.