rh56f1_hookonly_multicam (LeRobot v3 format)

LeRobot-format port of CrazyMoment/teleop_recorded_rh56f1_hookonly, keeping both cameras (d405 + zivid) for multi-view diffusion-policy training.

Robot	HDR35_20 (arm) + RH56F1_R (right dexterous hand)
Task	"remove hook ring from chassis"
Episodes	202
Frames	57,388
FPS	60
Cameras	d405 and zivid, both at 240 × 320 RGB
State dim	141
Action dim	12
Video codec	libsvtav1 (av1)
Codebase version	LeRobot v3.0

Differences from the source dataset

• The raw CrazyMoment/teleop_recorded_rh56f1_hookonly ships per-session HDF5 files (teleop_recorded_rh56f1_hookonly_<SESSION>.hdf5 + _images.hdf5) plus full-resolution videos. Total ~78 GB.
• This port flattens every (session, demo_i) pair into a LeRobot v3 episode, downsamples zivid from 1050×1458 to 240×320 (matching d405), and re-encodes both streams as libsvtav1 MP4 video. Total: ~0.9 GB.
• Only obs (141-d state) and actions (12-d) from the main HDF5 are kept; processed_actions and initial_state are dropped.
• zivid_mask and per-demo trajectory PNGs are dropped.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Ngseo/rh56f1_hookonly_multicam")
print(len(ds), ds.meta.features.keys())
sample = ds[0]
# sample["observation.state"]:        (141,)         float32
# sample["observation.images.d405"]:  (3, 240, 320)  float32 in [0,1]
# sample["observation.images.zivid"]: (3, 240, 320)  float32 in [0,1]
# sample["action"]:                   (12,)          float32

Reproducing this port

# raw HDF5 release → LeRobot v3 (multicam)
python examples/port_datasets/port_hookonly.py \
    --src   /path/to/teleop_recorded_rh56f1_hookonly \
    --out   /path/to/lerobot_data/rh56f1_hookonly_multicam \
    --robot rh56f1 \
    --cameras d405,zivid \
    --repo-id local/rh56f1_hookonly_multicam \
    --streaming-encoding         # ~6× faster than the default

Trained policy

A FlowMatch + DINOv2-S + multicam + DR Diffusion Policy trained on this dataset is available at Ngseo/rh56f1_diffusion_dinov2s_flowmatch_multicam_dr.

Acknowledgement

All source teleop demonstrations are © the authors of CrazyMoment/teleop_recorded_rh56f1_hookonly. This repo only redistributes a format-converted, downsampled, re-encoded subset.