Hugging Face's logo

Datasets:
sebastian-stapf
/
action-progress-droid

Modalities:
Video
Tags:
video
action-progress
segmentation
License:
Dataset card Files
xet
 Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

DROID is released under CC-BY 4.0; this repository is gated for consistency with the other action-progress sets. Access is granted manually.

Log in or Sign Up to review the conditions and access this dataset content.

Action progress: DROID (s31)

Training data of the droid_s31 arm of Whose Time Is It? Instance-Specific Action-Progress Control for Video Editing (LTX-2.5 actor-progress LoRA). Robot manipulation windows from DROID with the trajectory ground-truth progress.

Access. DROID is released under CC-BY 4.0; this repository is gated for consistency with the other action-progress sets. Access is granted manually.

Contents

file content
annotations.jsonl one record per training clip (fields below)
holdouts.jsonl the validation clips
masks/part-*.zip <clip>.npz: actors, objects uint8 [K, N, H, ceil(W/8)] (np.packbits over width), shape [H, W]; frame i is the clip's frame grid (see grid); <clip>.json: SAM 3.1 phrases, chosen instances, scores
videos/part-*.zip the native source video of every clip, byte for byte
configs/ base and training configs (LTX-2.5 22B, rank-32 LoRA, Gerard Donahue's RIFE-16 actor-progress method)

Canvas, frames and progress

  • canvas 960x544 (aspect kept, grey bars), 24 FPS, 97 frames per clip; frame grid = the native capture frames of the window (about 14 Hz); RIFE retimes them to 24 FPS
  • progress: ground truth: the episode-scale phase of the robot trajectory (phase), read at the native capture times
  • variants per clip: v00 natural, v01 forward warp, v02 rewind, v03 zigzag (warps that run backwards; profile extreme). Each variant shows the frames of a 16x RIFE interpolation of the window at dense_ids (dense frame 16k is grid frame k) and is conditioned on the measured progress at source_positions. The RIFE videos and VAE latents are not included; they are rebuilt from the source with the repository's preprocessing.
  • masks: SAM 3.1 from text (robot arm; objects from the language instruction). The training mask of an actor is its own mask united with its objects' masks.
  • captions: LTX-style captions written by Gemma-4-31B with the LTX trainer's caption instruction (ltx_trainer.captioning.DEFAULT_VIDEO_CAPTION_INSTRUCTION) from 8 frames per clip, plus the original annotated task.

Counts

3,193 training clips, 4 hold-outs, 3,197 source videos

Source and citation

DROID (Khazatsky et al., 2024), CC-BY 4.0. Masks, captions and windows are ours.
Downloads last month
7
Total file size:
16 GB