|
Download DATA_DICTIONARY.md from Bigenlight/cube_stack_raw: direct link, hf CLI and curl.
- Browser
- Download file 1.09 kB
-
https://huggingface.co/datasets/Bigenlight/cube_stack_raw/resolve/main/DATA_DICTIONARY.md
- Command line
-
hf download hf://datasets/Bigenlight/cube_stack_raw/DATA_DICTIONARY.md
-
curl -L -o DATA_DICTIONARY.md https://huggingface.co/datasets/Bigenlight/cube_stack_raw/resolve/main/DATA_DICTIONARY.md
1.09 kB
| # Raw recording dictionary | |
| Each `raw/<original-take-name>/` contains unchanged `vectors.h5`, `cam1.mp4`, and `cam2.mp4`. | |
| HDF5 groups contain one float64 dataset per named column. `ur_joint_states/q1..q6` are measured joints (radians); `command/cmd1..cmd6` are commanded joint targets (radians). `gripper/grip_pos` and `grip_cmd` use 0=open, 1=closed. GELLO leader coordinates are not policy observations. | |
| `cam1_frames` and `cam2_frames` map zero-based `frame_idx` to `stamp_s` (ROS header seconds) and `t_rel_s` (callback seconds relative to take start). Cam1 is the scene view; cam2 is the wrist view. The empty `synchronized` group is not used. | |
| The LeRobot conversion aligns camera/joint header stamps and command/gripper callback clocks using nearest timestamps. It retains all cam1 frames, with no servo-lag subtraction. State and action each contain six joints plus gripper. No depth was recorded. | |
| `source_takes.json` preserves collection tags and episode order; `checksums.json` records SHA-256 for every raw recording file. No success labels or paper evaluation split are asserted. | |