---
license: cc-by-4.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- manipulation
- synthetic-augmentation
- reactor-x2
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# reactor_x2_100

79 episodes, 6,818 frames of `so101_follower` arm data, each episode built by
adding a different distractor-fruit combo to one real recorded pick-and-place
episode with [Reactor](https://reactor.ai) XMAX X2 video editing. Standard
[LeRobot](https://github.com/huggingface/lerobot) v2.1 layout
(`meta/`, `data/`, `videos/` at the repo root), so it loads directly with:

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("kabilanKB/reactor_x2_100")
```

## Task

`"Grab orange and place into plate"` — every episode keeps the same recorded
grasp; only the scene around the target orange changes.

## Source data

- Base episode: `LightwheelAI/leisaac-pick-orange`, `episodes/orange_ep0`
- Robot: SO-101 follower arm, 6-DoF action / state, single front camera
  (480x640, 30 fps)

## How it was built

`build_fruit_addition_dataset.py` re-renders the source episode 95 times
(one per scene in `fruit_scenes95.json`), each time asking Reactor X2 to add a
different distractor fruit under a different lighting style, while the
original orange, plate, and arm motion stay put. Every output frame is gated
on the **recorded orange surviving per-frame** (not just a per-episode
average) — a scene only keeps the frames where the target is still
identifiable, and is truncated at the first sustained run of frames where it
isn't. This is the "add distractor objects" result from this project's
`PROJECT.md`: it works, conditionally, because the original action labels are
only valid where the target the policy was reaching for is still visibly the
same object.

Of 95 scenes attempted, **79 survived** with enough usable frames to keep;
16 were dropped for producing too few frames after the survival gate (see
`meta/dropped.json` — mostly heavy occlusion cases like `kiwi_lamp` and
`pear_plum_midday`, which kept 0 frames).

## Scene variation

- **9 distractor fruits**: apple, banana, grape, kiwi, lemon, peach, pear,
  plum, strawberry (single fruits and multi-fruit combinations)
- **5 lighting styles**: gold (warm late-afternoon), overcast, midday,
  lamp, morning

## Dataset stats

| | |
|---|---|
| Episodes | 79 (of 95 scenes attempted) |
| Frames | 6,818 |
| FPS | 30 |
| Action / state dims | 6 |
| Camera | `observation.images.front`, 480x640 |
| Mean per-frame coverage | 0.873 |
| Mean target ΔE (median / p90) | 20.7 / 24.8 |
| Episodes truncated by the survival gate | 10 / 79 |
| Frames kept vs. rendered | 6,818 / 7,108 (96%) |

Per-episode provenance (scene id, prompt, coverage, ΔE, truncation, source
frame range) is recorded in `meta/augmentations.jsonl`.

## Caveats

- This augments an existing action-labeled episode; X2 is pixels-only and
  contributes no actions, proprioception, or camera pose of its own.
- Whether training on this augmented data actually improves a policy is
  **unproven** — no A/B run against the un-augmented source has been done yet.
- Coverage and ΔE are automated proxies for "does the label still match the
  pixels," not a guarantee of visual quality.