madokalif/b1k-pick-up-from

BEHAVIOR-1K pick up from segments

Grasp segments: lifting a named object off a named surface.

Derived work. Every frame here comes from behavior-1k/2026-challenge-demos; this repository only re-cuts it into per-skill segments and adds the goal sentences. That dataset's license and terms govern this one — no separate license is claimed.

Skill segments cut from the BEHAVIOR-1K 2026 challenge demos, in LeRobot v3.0 layout. One annotated skill segment becomes one episode, and the goal sentence becomes that episode's task string — so prompt_from_task=True is the whole goal-conditioning mechanism.

episodes (segments)	9,703
frames	4,772,224 (44.2 h at 30 fps)
goal prompts	65
source tasks	16 of the 16 shared task files
robot	R1Pro, 23-dim action, 61-dim state

Example prompts: pick up allen wrench from countertop, pick up allen wrench from toolbox, pick up bottle of coffee from countertop

Videos are not included

videos/ is intentionally absent: the mp4 files are byte-identical to ones already public in behavior-1k/2026-challenge-demos, and meta/episodes/* references them by chunk/file/timestamp rather than copying them.

Everything needed to fetch them is in this repository, under source_files/. It used to point at a private code repo, which made this section unfollowable from outside; the list is now here.

file	what it is
source_files/files.txt	the exact 549 source paths — 360 mp4 + 186 parquet + meta
source_files/file_sizes.json	expected size of every file, so a partial download is caught
source_files/fetch_source_files.py	downloads exactly that list, and nothing else

The same 549 files serve all three sibling datasets — nav, pick and place were cut from the same 16 task directories and share one video tree, so fetch this once for all of them. Total 69.7 GiB (63.7 GiB mp4, 5.9 GiB parquet). huggingface-cli download --include cannot express the selection: it is a list of individual chunk/file paths across four trees, not a glob. Fetching by explicit path is also what keeps depth out — depth alone is 2,178 GB of the source repo's 3.26 TB and this pipeline never reads it.

# 1. fetch the source files (resumable; re-running skips what is already the right size)
python source_files/fetch_source_files.py \
    --files source_files/files.txt --dest /path/to/2026-challenge-demos

# 2. point this dataset at the video tree
ln -s /path/to/2026-challenge-demos/videos <this-dataset>/videos

The task directories covered are task-0000, task-0010, task-0012, task-0019, task-0052, task-0059, task-0060, task-0062, task-0063, task-0065, task-0069, task-0077, task-0090, task-0092, task-0093, task-0099.

Everything else — row data, episode metadata, per-episode video pointers, task strings, statistics — is here.

Splits

meta/episodes/* carries a split column and meta/val_episodes.json lists the held-out episodes (483 of them, stratified by source task and disjoint at the source-episode level). The split is recorded, not enforced: the openpi PyTorch trainer has no validation path and trains on every episode. It is there to be used by an evaluation that wants it.

How this was built

python build_nav_manifest.py --src <2026-challenge-demos> --out pick_manifest --skill "pick up from"
python build_nav_dataset.py  --src <2026-challenge-demos> \
    --manifest pick_subset/pick_up_from_subset.parquet --out b1k --name pick_up_from

The build scripts and the training recipe live in a private repository, so the pieces you need to USE this dataset are carried here instead: source_files/ reconstructs the videos, and meta/ carries the episode boundaries, video pointers and task strings. The π0.5 base these adapters train from is public at madokalif/pi05-b1k-base-pytorch-sft.

This is one member of a skill library: a shared π0.5 base with one LoRA adapter per skill. The sibling datasets are madokalif/b1k-nav-move-to, madokalif/b1k-place-in-on.