---
license: cc-by-sa-4.0
---

# Qiyuan Robotics Home Manipulation Challenge Dataset

> 中文版本：[README.md](README.md)

## UMI Sample Data

This directory is a standard **LeRobot v2.1** dataset containing 5 bimanual manipulation
episodes, together with left and right wrist-mounted ego-camera data.

```bash
pip install "lerobot==0.3.3"
```

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="local/umi_sample_data_v21",
    root="/path/to/umi_sample_data_v21",
)
```

### Overview

| Episode | `task_index` | Task | Frames | Duration |
| :--- | ---: | :--- | ---: | ---: |
| `episode_000000` | 0 | fold the red shirt | 1410 | 47 s |
| `episode_000001` | 1 | fold the black shirt | 1050 | 35 s |
| `episode_000002` | 2 | fold the yellow shirt | 870 | 29 s |
| `episode_000003` | 1 | fold the black shirt | 1050 | 35 s |
| `episode_000004` | 3 | fold the brown shirt | 1560 | 52 s |

The dataset has 5940 frames and 4 unique tasks, at 30 FPS, with 960 × 960 ego video.

### Directory structure

```text
umi_sample_data_v21/
├── data/chunk-000/                       # 5 episode Parquet files
├── videos/chunk-000/
│   ├── observation.images.left_ego/      # left-hand ego video
│   └── observation.images.right_ego/     # right-hand ego video
├── meta/
│   ├── info.json                         # dataset and field definitions
│   ├── tasks.jsonl                       # task to task_index mapping
│   ├── episodes.jsonl                    # episode lengths and tasks
│   ├── episodes_stats.jsonl              # per-episode statistics
│   └── calibration.json                  # camera and IMU calibration
├── annotation/                           # episode-level task and action-segment annotation
└── imu/                                  # left and right IMU data
```

### Field reference

#### Images

| Dataset field | Source |
| :--- | :--- |
| `observation.images.left_ego` | Left wrist-mounted RGB ego camera, 960 × 960 |
| `observation.images.right_ego` | Right wrist-mounted RGB ego camera, 960 × 960 |

#### Proprioception and actions

`observation.state` and `action` are both 16-dimensional, in the same field order:

| Indices | Field order | Meaning | Unit |
| :--- | :--- | :--- | :--- |
| 0–2 | `left_x, left_y, left_z` | Left end-effector position | m |
| 3–6 | `left_qw, left_qx, left_qy, left_qz` | Left end-effector quaternion `(w, x, y, z)` | - |
| 7 | `left_gripper` | Left gripper opening angle | ° |
| 8–10 | `right_x, right_y, right_z` | Right end-effector position | m |
| 11–14 | `right_qw, right_qx, right_qy, right_qz` | Right end-effector quaternion `(w, x, y, z)` | - |
| 15 | `right_gripper` | Right gripper opening angle | ° |

`observation.state` is the current-frame state. Except for the final frame,
`action[t] = state[t+1]`; the final frame's action keeps the next target from the original
capture sequence, so it is not necessarily equal to this episode's final state. Left and right
poses live in independent coordinate frames — you cannot directly compute the relative distance
or orientation between the two hands.

#### Index fields

| Field | Type | Description |
| :--- | :--- | :--- |
| `timestamp` | float32 | Time within the episode, in seconds |
| `frame_index` | int64 | Frame number within the episode, 0-based |
| `episode_index` | int64 | Episode number, range 0–4 |
| `index` | int64 | Global frame number in the dataset, range 0–5939 |
| `task_index` | int64 | Task number, keyed to `meta/tasks.jsonl` |

#### Annotations and IMU

| File | Description |
| :--- | :--- |
| `annotation/episode_subtasks_*.jsonl` | Episode-level task, target objects and success flag |
| `annotation/action_steps_*.jsonl` | Fine-grained action segments |
| `imu/episode_*_{left,right}.csv` | Left and right timestamps, three-axis angular velocity and acceleration |

Annotation intervals use the half-open convention `[start_frame_index, end_frame_index)` —
the start frame is included, the end frame is not.

The `subtask_index` field of `action_steps` is **constantly 0** across this delivery. It is a
placeholder — do not use it as a foreign key into `episode_subtasks`; derive the mapping between
the two layers from frame-interval overlap instead. This dataset carries exactly one
`episode_subtasks` row per episode (= the whole-episode task).