MemWAM / README.md
haodotgu's picture
Add packed index, calibration and README
9adf01d verified
|
Raw History Blame Contribute Delete
2.99 kB
---
license: mit
task_categories:
- robotics
tags:
- unitree-r1
- teleoperation
- manipulation
pretty_name: MemWAM
---
# MemWAM: Unitree R1 teleoperation recordings
Teleoperated manipulation episodes recorded on a Unitree R1 with its head
camera (egocentric view), joint state and Piper grippers.
| Task | Groups | Episodes | Duration | Sample images | Raw frames |
|---|---|---|---|---|---|
| `r1_order_retrieval` | 20 × 7 | 140 | 52 min | 31,242 | 46,106 |
| `r1_sorting` | 20 × 2 | 40 | 32 min | 18,808 | 27,755 |
| `r1_hidden_color` | 20 × 6 | 120 | 52 min | 30,605 | 45,187 |
Every image is provided twice: as recorded (fisheye) and undistorted (pinhole).
## Repository layout
```
packed/
<task>/<task>_000-019.tar one uncompressed tar per task (groups 000-019)
index.json per tar: groups, episodes, file counts, bytes, sha256
calibration/
cam_param.txt R1 head stereo factory calibration
undistort_r1.py script that produced the undistorted images
README.md
```
## Download
```bash
# one task
hf download haodotgu/MemWAM --repo-type dataset --include "packed/r1_sorting/*" --local-dir MemWAM
cd MemWAM && tar xf packed/r1_sorting/r1_sorting_000-019.tar
```
Use `--include "packed/*"` for everything. Each tar extracts to
`<task>_group_NNN/episode_N/`. Check integrity with the `sha256` in
`packed/index.json`.
## Episode format
```
episode_N/
episode.json metadata: task, times, frame counts, storage, provenance
samples.jsonl 10 Hz state rows (also as arrays in samples.npz)
samples.npz qpos/qvel (26), quat (4), token_state (64), gripper_* (2), t, camera_ego_t
images/ego/NNNNNN.png image matched to each sample row (10 Hz)
images/ego_undistorted/NNNNNN.png same, undistorted
frames/ego/NNNNNNNNN.png every camera frame received (~15 Hz)
frames/ego_undistorted/NNNNNNNNN.png same, undistorted
frames.jsonl timestamp and frame_id of every raw frame
controller.json controller / policy configuration
inference_meta.json
undistort.json calibration and geometry used for the undistorted images
```
Images are 384×216 RGB PNG (resized with INTER_AREA from the camera's
1280×720 stream). Sample row *i* uses `images/ego/{i:06d}.png`; raw frames are
indexed by `frames.jsonl`. Undistorted directories use the same file names.
Timestamps are Unix seconds.
## Camera and undistortion
The ego image is the R1 head stereo module's right-eye crop: rows 184–903 of
the 1280×1088 right sensor. Factory calibration (`calibration/cam_param.txt`)
uses the OpenCV rational model (k1 k2 p1 p2 k3 k4 k5 k6); the left/right
extrinsics give a 58.9 mm baseline.
Undistorted images are an ideal pinhole projection at 384×216 with
fx = fy = 114 px, principal point (191.5, 107.5), FOV 118.6° × 86.9°.
Areas outside the camera's view are black. The exact matrices are in each
episode's `undistort.json`.