README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- robotics
- retargeting
- agilex-piper
- lerobot
---

# piper-retargeted

Human demonstrations retargeted to an **AgileX Piper** arm with the Fidelity
Dynamics dt-pipeline (stage 6). Two datasets, one repo.

| archive | size | clips | frames | source |
|---|---|---|---|---|
| `piper_ego_retargeted.tar` | 2.2 GB | 1,574 | 164,601 | EgoDex-derived egocentric pick-and-place (`ego_filtered`, LeRobot v3.0, 30 tasks) |
| `piper_stera_plate_retargeted.tar` | 332 MB | 226 | 30,584 | stera-10m plate handling (`plate200_pp`) |

Each archive is **self-contained**: the observation video sits in the same clip
directory as the trajectory it belongs to, so there is nothing to join.

## Layout

```
<dataset>/clips/<clip_id>/observation.mp4   # RGB observation for this clip
<dataset>/clips/<clip_id>/retargeted.hdf5   # trajectories + per-frame IK error
<dataset>/clips/<clip_id>/report.json       # per-clip M1–M6 record, QA grades
<dataset>/manifest.json                     # aggregate stats + one row per clip
<dataset>/README.md
```

Video and trajectory are 1:1 per clip and share the clip's frame count and 30 fps
timebase. Archives are uncompressed `.tar` — the MP4s are already compressed, so
gzip would cost time and save nothing.

`retargeted.hdf5`:

| key | shape | meaning |
|---|---|---|
| `joint_positions` | (T, 14) | arms in R, L order — 6 joints + 1 gripper each |
| `ik_errors_R` / `_L` | (T,) | per-frame IK position error, metres |
| `base_R` / `_L` | (3,) | base placement chosen by BPP |
| `timestamps` | (T,) | seconds, 30 fps |

Attributes carry `joint_names`, `active_arms`, `parked_arms`, `arm_mode`,
`n_arms`, `fps` and the M4/QA records.

## Robot

`agilex_piper_bimanual`. Stage 6 loads one single-arm MJCF and drives it once
per arm with its own base placement, so "bimanual" needs no two-arm model.
Joints `joint1..joint6` plus jaw `joint7` (`joint8` is coupled `= -joint7`).
The TCP is the `grasp_frame` fingertip midpoint, not `Link6`.

Workspace placement used `WORKSPACE_CENTER_MODE=capmap`: the centre is the
reachability-index-weighted centroid of this arm's own pose-reachable voxels
(`[0.034, 0.000, 0.208]` over 18,479 voxels), not a reference robot's vector
rescaled. Scale 0.8811 (reach ratio 1.037; Piper 0.897 m vs reference 0.865 m).

## Quality

| | ego | stera |
|---|---|---|
| FK error (worst active arm) | mean **4.48 cm**, median 3.87, p90 7.63 | mean **10.53 cm**, median 7.20, p90 24.66 |
| saturation | mean 0.426 | mean 0.506 |
| QA PASS / WARN / FAIL | 262 / 1012 / 300 | 19 / 111 / 96 |
| dominant FAIL cause | saturation (331) | saturation (89), self-collision (18) |

**Filter, don't discard.** `qa_verdict` gates on the worst single arm joint's
fraction of frames pinned at a limit, so one saturated DOF fails a clip whose
tracking is otherwise good. `manifest.json` carries per-clip `ik_mean_cm` and
`saturation` if you would rather threshold yourself. The stera set is
substantially the weaker of the two.

### Per-dataset notes

**ego** — both arms are active on every clip, but not equally busy: the right
arm travels a median 1.00 m per clip against the left's 0.025 m and carries the
error (4.25 vs 0.44 cm). A near-static arm can still saturate by resting against
a limit; 139 of the 300 FAILs are left-arm.

**stera** — single-arm (`--arm right`); the left arm is a constant park pose
carrying no signal, so drop columns 7–13 for single-arm training. **29 of the
255 selected clips are absent** because the human demonstrator used their left
hand, dropping the right below the 50% tracking threshold; they are listed in
`manifest.json` under `skipped_clips`. Because M3.5 re-solves the moving arm to
avoid the parked one, 18 clips FAIL on `selfcol` for a collision that will not
exist once the left arm is dropped.

## Known limit

About 73% of Piper's joint saturation is **wrist-driven** (joint4 and joint5;
joint5's range is only ±1.22 rad, the tightest on the arm). A workspace
re-placement cannot fix an orientation the wrist cannot reach, which is why the
teleop-supervised refit is disabled for these runs — measured, it accepted 2 of
150 clips. The remaining lever is orientation-side, not placement.