---
license: cc-by-nc-4.0
task_categories:
- robotics
tags:
- robotics
- manipulation
- lerobot
- droid
- physical-intelligence
- imitation-learning
- franka
size_categories:
- 10K<n<100K
pretty_name: MatrixNorm RoboticsGraph · DROID (free 327-episode sample)
---

# RoboticsGraph · DROID — FREE SAMPLE (327 episodes, videos included)

**This is a free sample of a much larger commercial release.** The full version —
**1,700,000 frames / 5,819 episodes / 70 collection sites**, with a
13.7M-edge relationship graph and full video linkage — is available for
purchase. Contact: **energea@energeatech.com**.

**MatrixNorm RoboticsGraph** turns raw robot-learning corpora into
training-ready, rights-tracked, leakage-safe datasets: the verified chain
`instruction → observation → state → action → outcome` as structured rows —
not raw robot video dumps.

## What's in the sample

327 real-world Franka manipulation episodes (96,183 timesteps, 45 collection
sites) — and **every sample episode's exterior-camera video is included and
frame-addressable** from the data rows.

| file | what it is |
|---|---|
| `roboticsgraph_dataset.jsonl` / `.csv` | THE flat normalized table — one row per timestep, 38 columns |
| `training_tasks.jsonl.gz` | 96,183 training tasks: instruction + full robot_state → action vector |
| `train.jsonl.gz` / `validation.jsonl.gz` / `test.jsonl.gz` | the same tasks, materialized by episode-grouped split (220/47/60 episodes — zero episode leakage) |
| `eval_tasks.jsonl` | 327 held-out `predict_success` eval tasks |
| `hin_edges.jsonl.gz` | 774,067-edge relationship graph (episode/timestep/state/action/outcome/task) |
| `videos/` + `videos_index.*` | the covering exterior-camera MP4 + per-episode/per-camera windows; `video_frame_timestamp` on every row seeks straight to the frame |
| `source_manifest.json` / `rights_manifest.csv` | provenance, license trail, raw-archive SHA256 |

```python
from datasets import load_dataset
rows  = load_dataset("json", data_files="roboticsgraph_dataset.jsonl", split="train")
tasks = load_dataset("json", data_files="train.jsonl.gz", split="train")
```

## Field coverage (this sample)

- `robot_embodiment` (Franka, verified from source metadata): **100%**
- state / end-effector pose / gripper / action vectors: **100%**
- video linkage (file + exact frame timestamp per row): **100%**
- collection site + collector provenance: **100%**
- outcome (success/failure ground truth): **100%**
- language instruction: **72.8%** — the remainder have no label anywhere in the
  source; they are flagged, never invented, and their tasks are typed
  `observation_to_action`. Alternate source phrasings are preserved.

## Why this beats using the raw source directly

1. **Training-ready tasks** — every timestep is already an
   (instruction, robot_state, video-frame ref) → action_vector pair; no
   parsing, joining, or episode reconstruction needed.
2. **Leakage-safe splits** — episode-grouped, hash-materialized, auditable via
   `split_group_key`. No silent train/test contamination.
3. **Honesty contract** — every label traces to a real source field; gaps are
   flagged, not filled. Two corrupt source columns were detected and excluded
   (documented below).
4. **Rights tracked end-to-end** — apache-2.0 verified at the source, rights
   manifest + raw-archive SHA256 included.
5. **Relationship graph** — HIN edges connect timesteps, states, actions,
   outcomes and tasks for graph-based learning and retrieval.

## The full release (for sale)

- **1,700,000 frames / 5,819 episodes / 70 sites** (≈18× this sample)
- 1.7M training tasks + 5,819 eval tasks, 13,682,009 HIN edges
- full video index for all episodes × 3 cameras (exterior 1/2 + wrist)
- same validated 100% coverage on embodiment/state/video/provenance fields
- custom normalization runs (other robots, other corpora, up to the full
  95,658-episode DROID) available on request

**Contact: energea@energeatech.com**

## Honesty notes

- The source mirror's `task_category` and `date` columns are corrupt duplicates
  of `building` and `collector_id` (verified) — excluded rather than shipped as
  fake labels.
- `gripper_action_state` open/closed is a disclosed 0.5 threshold over the
  continuous `gripper_position` (preserved verbatim).
- Instructions are never synthesized; unlabeled episodes stay unlabeled.

## License

This packaged sample (the flat table, task construction, splits, HIN graph,
video index, and this documentation) is licensed **CC BY-NC 4.0** —
attribution required, **non-commercial use only**. Commercial use requires a
license — contact **energea@energeatech.com** (the full 1.7M-frame release is
sold under a separate commercial license).

**Scope note:** this NC restriction covers williamTLmiller's normalization and
packaging work. It does not and cannot restrict the underlying DROID
recordings themselves, which remain separately available under Apache
License 2.0 from their original source
([`lerobot/droid_1.0.1`](https://huggingface.co/datasets/lerobot/droid_1.0.1)).
See `LICENSE` (CC BY-NC 4.0, this package) and
`LICENSE-APACHE-2.0-UNDERLYING-DATA.txt` (the source data's own license).

## Attribution

Derived from [DROID](https://droid-dataset.github.io/) via the official
LeRobot-format mirror [`lerobot/droid_1.0.1`](https://huggingface.co/datasets/lerobot/droid_1.0.1)
(Apache-2.0). Please cite the DROID paper when using the underlying data.