skysky0214's picture
Add v01 dataset
ddbda0d verified
# evButtonPush-260422-01-isaaclab
Call-button pressing demonstrations for OMY (6-DoF arm + 2-finger gripper) in Isaac Lab, collected with a scripted IK teacher (distance-threshold based).
## Summary
- **Task**: `RobotisLab-CallButton-Right-OMY-v0`
- **Robot**: OMY on a 0.908 m pedestal
- **Demos**: 46 (all successful, no failures)
- **Format**: IsaacLab HDF5 (`ActionStateRecorderManagerCfg`)
- **Collection dates**: 2026-04-22 ~ 2026-04-23
- **Total size**: ~89 GB
## Environment
- Task ID: `RobotisLab-CallButton-Right-OMY-v0`
- Elevator USD: random choice from 100 variants (`elevator_setup_new`)
- Robot randomization: standoff ∈ (0.55, 0.65), lateral ∈ (−0.05, 0.05), yaw=0
- Gripper: always closed (stiffness 2e6, effort limit 5000)
- Pedestal height: 0.908 m
## Cameras (all 1920×1200 RGB)
| Name | Prim path | Frame | Position (m) | Quat (w,x,y,z) |
|---|---|---|---|---|
| `cam_wrist` | `Robot/OMY/link6/cam_wrist` | link6 (local) | (0.000, −0.080, 0.070) | (0.0018, −0.0018, 0.7071, 0.7071) |
| `cam_top` | `Robot/OMY/world/cam_top` | robot base | (0.000233, 0.1499, 0.43953) | (0.5144, 0.4146, −0.5064, −0.5542) |
| `cam_belly` | `Robot/OMY/world/cam_belly` | robot base | (0.08705, 0.1499, 0.1075) | (−0.4901, −0.5097, 0.5097, 0.4901) |
All use Pinhole: focal 10~11.8 mm, horizontal aperture 20.955, clip (0.01, 100).
## Teacher (distance-based success)
- Script: `diffik_teacher_jointpos.py` (older version, no force-contact)
- Success criterion: `min_distance ≤ contact_threshold (0.05 m)`
- No elevator door interaction, no LED observation
### CLI parameters
```
--approach-steps 100
--press-steps 400 (earlier batches used 300)
--stall-steps 15
--contact-threshold 0.050
--hold-steps 30
--return-steps 150
--press-depth 0.010 (or 0.015 for some)
```
## Z-offset subsets
Teacher was iteratively tuned — pressing height shifted over three batches:
| Z offset | Seeds | Count | Note |
|---|---|---|---|
| 0.00 m | 00–32, 43, 44 | 35 | Original teacher, press target at button center (user deemed ~1 cm too low) |
| +0.01 m | 50–54 | 5 | Mid-batch adjustment, still ~1 cm low |
| +0.04 m | 55–60 | 6 | Teacher `button_base[2] += 0.04` hack, pressed 4 cm higher |
Missing seeds (33–42, 45–49, 61–64) were interrupted or skipped.
## Recording contents (per HDF5)
Each `demo_XX.hdf5` contains a single `data/demo_0` group with:
- `actions`, `processed_actions`
- `obs/cam_wrist`, `obs/cam_top`, `obs/cam_belly` (uint8 RGB)
- `obs/joint_pos`, `obs/joint_vel`, `obs/actions`
- `obs/call_button_pos`, `obs/rel_ee_call_button_distance`
- `states/articulation/robot/*`, `states/rigid_object/pedestal/*`
- `initial_state/*`
No `call_button_lit` or force observations (those are in v02 dataset).
## Files
- `demo_XX.hdf5` — per-episode recording
- `demo_XX.log` — teacher stdout (includes [SETUP], approach/press step traces, [RESULT])
- `summary.log` — one-line per-demo result
- `metadata.csv` — structured per-demo metadata
## Known limitations
- Heterogeneous press target height (3 Z offsets) — inconsistent for end-to-end imitation
- No physical contact detection; pressing success judged only by distance
- No elevator door / LED state — pure static-scene press
- Recommended use: **baseline / ablation**. For interaction learning, see `evButtonPush-260423-02-isaaclab` (v02).