mahgoobi/b1k_2026_extras

BEHAVIOR-1K 2026 challenge extras

Precomputed tables and preview clips derived from behavior-1k/2026-challenge-demos. The source release is 3.0 TB across 17,093 video shards; this is 111 MB, and it is enough to read the whole annotation layer and watch one demonstration per task without touching the release at all.

Contents

path	what it is
b1k_skill_segments.parquet	one row per skill span, all 20,000 episodes: 407,054 rows of episode_index, task_index, skill_idx, skill_id, skill_description, skill_type, start, stop, n_frames, n_objects
b1k_primitive_segments.parquet	the same for the primitive layer, which exists only on the 50 tasks carried over from the 2025 challenge: 95,199 rows
b1k_annotation_coverage.parquet	one row per episode: length, valid window, segment and span counts, covered frames, multi-labelled frames, interior gaps, head and tail
b1k_annotations/	the release's own annotation JSON for the 100 episodes the clips were built from
b1k_clips/	six cameras × 100 tasks, one representative episode each, sampled to at most 360 frames and re-encoded to browser-playable H.264
b1k_frames/	per-frame reward, cumulative reward, termination and the full skill record for each clip, aligned frame for frame
b1k_action_corr.npz	the 690×690 covariance of a 30-step action chunk (23 dims × 30 steps), accumulated over the corpus

What the annotation layer says

Measured over all 20,000 annotation files, not a sample:

• 35 skill verbs label 98.86% of the release's 210,916,774 frames, in 406,341 segments
• only 3,325 of 20,000 episodes are labelled end to end; the rest lose frames to a head, an interior gap, or a tail
• 0.66 M frames fall in interior gaps, 1.47 M in unlabelled tails, and 0.32 M carry two labels at once
• skill_type is one of navigation (38.1% of labelled frames), uncoordinated (52.6%) and coordinated (9.3%)
• memory_prefix is non-empty on 72,808 records, spatial_prefix on 43,131, and mp_ef on none

Clips

b1k_clips/task-NNN.mp4 is the head RGB camera; task-NNN-{left,right}-rgb.mp4 and task-NNN-{head,left,right}-depth.mp4 are the other five. All six are sampled to the same frames of the same episode, so one timeline drives them together.

Depth in the source release is not stored in metres. It is log-quantized over [0, 10] m with a 3.5 shift, and the 2026 files are gray12le, so codes run 0..4095:

import numpy as np
SHIFT, MAX_D = 3.5, 10.0
lo, hi = np.log(SHIFT), np.log(MAX_D + SHIFT)
metres = np.exp(code / 4095 * (hi - lo) + lo) - SHIFT

Decoding a 2026 file with the 14-bit qmax of the 2025-era loader overshoots by 4× and pins about two thirds of a frame at 10 m. The depth clips here are colour-mapped on that fixed scale (square-rooted so the near metre a wrist camera lives in is legible), not normalised per clip, so one colour is one distance in every clip.

Use

hf download mahgoobi/b1k_2026_extras --repo-type dataset --local-dir "$ROBORESEARCH_ARTIFACTS"

The layout matches what the BEHAVIOR-1K notebook in EAI-RSM/RoboResearch expects under $ROBORESEARCH_ARTIFACTS, so no renaming is needed.

Licence

MIT, following the source release. Everything here is derived from behavior-1k/2026-challenge-demos; cite BEHAVIOR-1K for the underlying data.