ewinchell07's picture
Add README.md
51eff65 verified
|
Raw History Blame Contribute Delete
3.31 kB
---
task_categories:
- robotics
tags:
- LeRobot
- anvil
- openyam
- teleoperation
- rustic-bakery
configs:
- config_name: default
data_files: data/*/*.parquet
license: mit
---
# rotary-table-flatbread
LeRobotDataset v3.0 of an OpenYAM (Anvil workcell) robot teleoperated with a Quest 3S at Rustic Bakery
(onsite, 2026-11-08) to **"pack sourdough flatbread bags from the rotary table into the case"** (single task).
> The task prompt was written from the video during processing; no prompt was recorded with the data.
| | |
|---|---|
| robot_type | `anvil_yam` |
| episodes / frames | 71 / 163,210 (45.3 min) |
| fps | 60 |
| cameras | `chest` (head), `wrist_l`, `wrist_r` — 480x270, 180° fisheye, h264 |
| observation.state / action | 14 = per arm `[finger_joint1, joint1..joint6]`, left then right; absolute joint positions |
**Single-arm data.** The left arm does the task; the right arm stays parked (≈0.05 rad of total motion, gripper
used in 5 of 63 raw episodes). Its state and action columns are present but nearly constant. The sister Lundberg
dataset (`titus-robots/rotary-table-rice-cakes`) is bimanual and right-dominant.
## Sessions
Two recording sessions merged with `lerobot` 0.5.1 `merge_datasets` (frames verified identical to the per-session
conversions). `anvil_qc/episode_map.csv` maps every episode to its session, segment and raw episode.
| episode_index | Session | Raw | Episodes here |
|---|---|---|---|
| 0–20 | `rustic-bakery-onsite-rotary` | 18 eps, 17.5 min | 21 (12.3 min) |
| 21–70 | `rustic-onsite-rotary-2` | 45 eps, 47.4 min | 50 (33.1 min) |
## Provenance
Raw MCAP: R2 `titus-datasets/captures/2026-11-08_anvil_rustic-bakery-onsite-rotary/` and
`…/2026-11-08_anvil_rustic-onsite-rotary-2/` (originally HF `JSklar03/rustic-onsite-2026-11-08`, sha256-verified).
Converted with `anvil-embodied-ai` `mcap-convert` (commit 5d1c8eb, config `anvil_qc/openyam_bimanual_quest_16x9.yaml`).
**Cleaning** (conservative; `anvil_qc/trim_teleop_gaps.py`, logs in `anvil_qc/<session>/trim_manifest.json`):
- Quest teleop dropouts (both arms' commands absent > 3 s) removed by **splitting** the episode at the gap
(0.5 s of stillness kept on each side), never spliced. Single-arm pauses kept.
- Idle head/tail > 3 s trimmed to 1 s; pieces < 5 s created by splitting dropped.
- Dropped: session 1 raw 0002 (no teleop).
- Skipped as `mcap-valid` critical (session 2): 0039 (0.5–0.8 s stall on all three cameras), 0040–0042 (right-arm
command topic absent for the whole episode). All four remain in the raw capture.
## QC
- `mcap-valid`: session 1 3 pass / 18 warning / 0 critical; session 2 5 pass / 45 warning / 3 critical (skipped).
Warnings are single-arm idle topics.
- `dataset-valid`: passed. Action columns track their matching joints; no NaNs. Debug plots per session in `anvil_qc/`.
- **Image quality is uneven.** Session 1 `chest` is hazy and low-contrast (Laplacian sharpness ~160 vs ~1,500 at
Lundberg). In session 2 `chest` improves (~250 → ~1,350) while `wrist_l` softens towards the end (~70–100).
- Session 2 has short all-camera stalls (> 200 ms) in 34 of 45 raw episodes (workcell load), max 1.8 s.
- Per-episode raw QC JSON in `anvil_qc/<session>/per_episode_qc_raw/`; comparison with Lundberg in
`anvil_qc/rustic_vs_lundberg.md`.