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

# piper-single-arm-cotrain

Five **LeRobot v2.1** datasets on an **AgileX Piper** arm, put on one schema so
they can be mixed without a rename map: two built by retargeting human
demonstrations, three from real teleoperation. Four form a duration-balanced
training mixture; the fifth is a held-out validation set.

### Training mixture

| archive | size | episodes | frames | duration | tasks | provenance |
|---|---|---|---|---|---|---|
| `piper_ego.tar` | 172 MB | 1,574 | 109,730 | 1 h 31 m | 30 | EgoDex-derived egocentric pick-and-place, retargeted |
| `piper_stera_plate.tar` | 75 MB | 226 | 20,385 | 17 m | 127 | stera-10m plate handling, retargeted |
| `mrfood1.tar` | 1.62 GB | 294 | 61,709 | 51 m | 1 | real teleop |
| `mrfood3.tar` | 1.74 GB | 285 | 68,271 | 57 m | 1 | real teleop |

Retargeted **130,115** frames vs teleop **129,980** — 1 h 48 m per side, 135
frames apart.

### Validation

| archive | size | episodes | frames | duration | tasks | provenance |
|---|---|---|---|---|---|---|
| `mrfood2.tar` | 1.85 GB | 347 | 67,991 | 57 m | 1 | real teleop, held out |

`mrfood2` is a **separate teleop recording session** from the two in the
training mixture (2026-08-19, against 08-14/15 and 08-11), carrying the same
schema and the same transforms — but **untrimmed**, since a validation set
should not be truncated to a training budget.

Read what it measures narrowly: it is the same task (`pick_and_lift_right`) on
the same rig, so it scores **cross-session generalisation within the teleop
domain**. Its single task is exactly the one the teleop training half already
covers. It says nothing about the retargeted domain, which contributes 157 of
the mixture's 158 distinct task strings (the two halves share no task) and all
of its synthesized imagery. A model that
scores well here has not been shown to transfer to the retargeted distribution,
or to any unseen task.

## Schema

Every archive — training and validation alike — is a self-contained LeRobot
v2.1 dataset with an identical schema:

```
<dataset>/data/chunk-000/episode_NNNNNN.parquet
<dataset>/videos/chunk-000/observation.images.top/episode_NNNNNN.mp4
<dataset>/videos/chunk-000/observation.images.right-arm/episode_NNNNNN.mp4
<dataset>/meta/{info,episodes,tasks,episodes_stats}.jsonl + info.json
```

| | |
|---|---|
| `codebase_version` | `v2.1` |
| `robot_type` | `agilex_piper` (single arm) |
| rate | 20 Hz / 20 fps |
| `observation.state`, `action` | `float32[7]` = `[j1…j6 in degrees, gripper ∈ {0,1}]` |
| cameras | `observation.images.top`, `observation.images.right-arm` |

Gripper is `0` = closed, `1` = open. Joint angles are **degrees**, within the
Piper limits (±150, 0–180, −155–0, ±105, ±70, ±180).

`meta/source_episodes.jsonl` maps each episode back to its index in the
pre-trim dataset; `meta/source_clips.jsonl` (retargeted) and
`meta/episode_quality.jsonl` (teleop) are carried through where the source had
them.

## How these were unified

The sources did not agree, in ways that raise no error at load time:

- **Left arm dropped.** It was a constant park pose in the retargeted sets and
  exactly zero in the teleop sets — no signal either way. This also moves the
  gripper to index 6 everywhere; it had been at index 6 in the retargeted sets
  and index **12** in teleop.
- **Radians → degrees** on the retargeted sets. Teleop was already degrees.
- **Gripper binarized.** Teleop was already effectively binary (one transition
  per clip), so a plain 0.4 threshold applies. The retargeted gripper is a
  continuous hand-aperture signal that never reaches either stop; a plain
  threshold on it yields a median of 3–4 transitions per clip — phantom
  grasp/release events inside a single pick-and-place. It instead gets per-clip
  min/max normalisation and a Schmitt trigger (0.40 ±0.15, 8-frame dwell),
  giving a median of 2 transitions (grasp + release).
- **20 Hz.** The retargeted sets were resampled from 30 Hz by linear
  interpolation onto a uniform 50 ms grid (30→20 is a 2/3 ratio, so decimation
  would leave alternating 33/67 ms gaps), with the action recomputed afterwards.

Gripper open-fraction after binarization: 0.64 / 0.49 (retargeted) against
0.56 / 0.51 (teleop).

## Known limits

**The retargeted views are synthetic, and share one camera.** `top` and
`right-arm` are both crops of a *single* egocentric frame — `top` is the square
centre crop, `right-arm` a zoom window tracking a MediaPipe-detected grasp
point. They carry no parallax, unlike the teleop sets' two physical cameras.
Video is 224×224 upright for the retargeted sets and 480×640 portrait
(rotated ~90°) for teleop.

**Action means slightly different things.** The retargeted action is exactly
`state[t+1]` — a perfect next-frame target, because it is constructed that way.
The teleop action is a separately recorded controller command that leads the
state by one frame with a real tracking residual (~0.08°). Both are absolute
next-step position targets.

**Joint envelopes only partly overlap.** The retargeted trajectories reach every
Piper joint limit; the teleop trajectories occupy a narrow interior band. About
73% of the retargeted saturation is wrist-driven (joint4/joint5; joint5's range
is only ±1.22 rad, the tightest on the arm).

**`piper_stera_plate` is the weakest set.** Its retargeting FK error is
substantially higher than `piper_ego`'s, and 10% of its clips still show more
than four gripper transitions after debouncing.

**Grasp events are inferred, not labelled.** The retargeted gripper signal comes
from hand pose, not from a recorded gripper command.

## Provenance and attribution

- `piper_ego`, `piper_stera_plate` — retargeted with the Fidelity Dynamics
  dt-pipeline (stage 6) from
  [`angkul07/piper-retargeted`](https://huggingface.co/datasets/angkul07/piper-retargeted),
  itself derived from EgoDex-style egocentric recordings and stera-10m.
- `mrfood1` — from [`PranayTest/pick-right-2026-08-14to15-2cam`](https://huggingface.co/datasets/PranayTest/pick-right-2026-08-14to15-2cam)
- `mrfood3` — from [`PranayTest/pick-right-2026-08-11-2cam-prxxx`](https://huggingface.co/datasets/PranayTest/pick-right-2026-08-11-2cam-prxxx)
- `mrfood2` — from [`PranayTest/pick-right-2026-08-19-2cam`](https://huggingface.co/datasets/PranayTest/pick-right-2026-08-19-2cam)

The teleop sources are LeRobot v3.0 and were converted to v2.1 here (v3.0 packs
many episodes per parquet and concatenates episodes into shared video files;
splitting those requires a re-encode, since the cut points are not
keyframe-aligned). `mrfood1` and `mrfood3` were then trimmed by dropping whole
episodes in a golden-ratio order — so the cut spreads evenly across each run
rather than concentrating on one stretch of a session — to match the retargeted
duration. `mrfood2` is untrimmed.

All upstream datasets are Apache-2.0.

## Validation

Every episode in all five archives passes: parquet row count against declared
length, contiguous `frame_index`, timestamps against the declared rate,
globally continuous `index`, and a frame count for each declared camera key.
Zero errors.

## Usage

```bash
tar xf piper_ego.tar
```

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("piper_ego", root="piper_ego")
```