---
license: cc-by-4.0
task_categories:
  - robotics
tags:
  - robot-manipulation
  - point-tracking
  - 3d
  - multi-view
  - subtask-segmentation
pretty_name: DROID subtask split — per-subtask temporal segmentation of DROID 3D point-track episodes
---

# DROID subtask split

`droid/subtask_split/` is a **per-subtask temporal split** of the DROID 3D point-track episodes. Each DROID episode is cut into its constituent subtasks along the time axis, so one episode becomes an ordered sequence of subtask clips — each with its own natural-language instruction, primitive-skill label, and its own temporally-sliced 3D point tracks, video, and calibration. Every subtask clip is a complete, self-contained sample in the **exact same format** as a normal DROID episode.

## Coverage

- **28,308 episodes**, **89,312 subtasks** (avg **3.16** subtasks/episode; per-episode subtask counts range 1–39, mode 3).
- **12 tar shards** `subtask_split-000.tar … subtask_split-011.tar`, ~327 GB total (eleven ~29.5 GB shards + a ~2.4 GB tail). All shards are under HF's 50 GB limit, so each extracts directly — no `.part` reassembly.
- **15 primitive skills** (subtask count): `pick` 26,195 · `transfer` 25,000 · `place` 24,379 · `push` 3,116 · `press` 2,356 · `pull` 2,045 · `slide` 1,253 · `fold` 1,183 · `manipulate` 1,160 · `pour` 1,023 · `throw` 698 · `twist` 672 · `insert` 153 · `shake` 57 · `strike` 22.

`manifest.json` maps every episode → shard and lists each episode's per-subtask `frame_range`, `primitive_skill`, subtask instruction text, episode-level `instruction_add`, and RoboInter provenance — so you can index the whole split without untarring.

## Layout

Each episode dir holds one subdir per subtask (mirrors `robocasa/subtask_split/`; only the `subtask_XX/` folders are shipped, not the un-split episode files):

```
<episode_id>/
  subtask_00/   view1.mp4  view2.mp4  robot.npz  gripper.npz  meta.json
  subtask_01/   ...
  ...
```

`<episode_id>` is the DROID id (`<lab>+<hash>+<timestamp>`). Every `subtask_XX/` is a complete sample in the standard format:

- `view1.mp4`, `view2.mp4` — the two exterior ZED views, re-encoded per subtask (libx264, **crf 18**, visually lossless).
- `robot.npz` — `points`: `(T, 512, 3)` float32, Franka Panda arm (non-gripper) CAD surface points from joint FK.
- `gripper.npz` — `points`: `(T, 256, 3)` float32, Robotiq 2F-85 CAD surface points.
- **All 3D points are in the `view1` camera frame** (OpenCV convention: +x right, +y down, +z forward; meters), identical to the parent dataset. `world_frame` is the DROID robot base; `T_cam_world` maps base-frame points into the camera.
- `meta.json` carries the standard DROID calibration (per-view `intrinsics` and `T_cam_world` 4×4 with `extrinsics_convention: "cam_from_world"`), plus a **`subtask`** block: `subtask_idx`, `n_subtasks`, `lang` (the subtask instruction), `primitive_skill`, `frame_range` (`[start, end)` into the original episode), `instruction_add` (the whole-episode instruction), `orig_id`, `orig_n_frames`, and RoboInter provenance (`roboint_index`, `roboint_key`, `annotation_source`). The top-level `meta.json → lang` is the subtask instruction and `n_frames` is the subtask length `T`. DROID cameras are a static rig, so the single `T_cam_world` per view is exact at every frame.

## Lossless reconstruction

The subtasks tile the original episode's `[0, n_frames)` contiguously — no gaps, no overlap. Concatenating an episode's `subtask_00 … subtask_k` in order **exactly reconstructs the original**: the `.npz` point arrays are **byte-identical** (`np.array_equal`) to the un-split episode, and `Σ subtask n_frames == original n_frames`. This was verified per-episode at build time (a build marker was written only after the concatenation matched the original byte-for-byte). Videos are re-encoded per subtask at crf 18 (visually lossless; the point tracks + calibration are the exact, bit-preserved data).

Rebuild an original episode's tracks:

```python
import numpy as np, glob
sds = sorted(glob.glob(f"{ep}/subtask_*"))
robot = np.concatenate([np.load(f"{s}/robot.npz")["points"] for s in sds])   # == original robot.npz
gripper = np.concatenate([np.load(f"{s}/gripper.npz")["points"] for s in sds]) # == original gripper.npz
```

## Annotation source

The per-subtask segmentation (time boundaries, subtask instruction, primitive-skill label) comes from **[InternRobotics/RoboInter-Data](https://huggingface.co/datasets/InternRobotics/RoboInter-Data)** (InternRobotics; **MIT license**), `Annotation_pure`. Please cite RoboInter-Data if you use these subtask annotations. Each subtask's `meta.json → subtask` records the exact `roboint_index` and `roboint_key` it came from.

**Join key:** annotations are joined 1:1 to DROID episodes on the DROID **lab + timestamp** (i.e. the episode id).

## Base data & licenses

The underlying episodes, video, calibration, and 3D point tracks are the DROID split of this dataset (derived from [DROID](https://droid-dataset.github.io/), CC-BY-4.0). The subtask segmentation adds the RoboInter-Data annotations (MIT) on top. See the top-level `README.md` for the full point-track format and the visualizer (`tools/visualize_episode.py`).