abcdl_demo

abcdl_demo is a bimanual robot-manipulation imitation-learning dataset (20 episodes, 71877 frames at 30.0 FPS) collected on a two-arm station with 4 cameras. Each frame carries a 14-D proprioceptive state and a 14-D action, plus synchronized camera images.

Published with abcdl — the ABC A Behavior Cloning Dataloader format. The same data is offered in several formats, each stored as its own git branch; every version is a tag <format>-<version> so resolutions and revisions coexist in one repo.

• mcap — full-resolution source recordings (the only branch you can re-process into other resolutions/alignments; everything else is derived from it).
• abcdl_<size> — the training cache: cameras stacked into one MP4 at <size>px plus a raw state/action binary, encoded for near-free random frame access.
• lerobot — a LeRobot v3.0 dataset (parquet + per-camera videos).

Available formats (branches)

branch	contents
lerobot	LeRobot v3.0 dataset (parquet + videos)
abcdl_224	abcdl training cache (MP4+binary) @ 224px
mcap	full-resolution source recordings (MCAP; the only re-derivable original)

abcdl_224 @ v1

field	value
Episodes	20
Frames	71877
Cameras	top_left, top_right, left_wrist, right_wrist
Resolution	224x224
FPS	30.0
State dim	14
Action dim	14
Robot	yam_bimanual

Tasks

• arrange the flowers into the vase

Install

pip install git+https://github.com/jellyho/abcdl_RLLAB    # provides the `abcdl` package

Usage (LeRobot-compatible loader)

from abcdl.dataset import AbcdlDataset
from torch.utils.data import DataLoader

# auto-downloads the `abcdl_224` branch from the Hub, then loads locally
ds = AbcdlDataset("jellyho/abcdl_demo", fmt="abcdl_224", version="v1")
print(len(ds), "frames |", ds.meta.camera_keys)

item = ds[0]
#   item["observation.state"]            -> FloatTensor (14,)
#   item["action"]                       -> FloatTensor (14,)
#   item["observation.images.<cam>"]     -> FloatTensor (3, H, W) in [0, 1]
#   item["task"]                         -> str

# action chunks (e.g. 16 steps) for chunked policies:
ds = AbcdlDataset("jellyho/abcdl_demo", fmt="abcdl_224", delta_timestamps={"action": [i/ 30.0 for i in range(16)]})

# standard multi-worker training loop:
loader = DataLoader(ds, batch_size=64, num_workers=8, shuffle=True)
for batch in loader:
    ...

Pull the raw full-resolution source instead:

from abcdl import hf
mcap_dir = hf.pull("jellyho/abcdl_demo", fmt="mcap", version="latest")

_{Dataset card auto-generated by abcdl.hf.push.}