---
title: RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3
canonical_url: "https://www.modelscope.cn/datasets/tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3"
md_url: "https://www.modelscope.cn/datasets/tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3.md"
repository: tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3
last_updated: 2026-10-19
license: Apache-2.0
storage_size: "7.1 GB"
downloads: 604
stars: 0
---

# RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3

> RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3 - tpy001 在 ModelScope 开源的数据集。RoboTwin Clean LeRobot v3: 2500 episodes, native 640x480 RGB and SAM-filtered global step-8 flow, with auxiliary videos

tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3 是 ModelScope 魔搭社区上的数据集，存储大小 7.1 GB，采用 Apache-2.0 许可。

- **Repository**: tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3
- **License**: Apache-2.0
- **Storage size**: 7.1 GB
- **Downloads**: 604
- **Stars**: 0
- **Last updated**: 2026-10-19

Source: https://www.modelscope.cn/datasets/tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3

---

# RoboTwin RGB + global step-8 optical-flow dataset

LeRobot v3.0, 2,500 episodes, 549,787 frames, 15 FPS, 640×480.

The original RGB, state, action, tasks and episode boundaries are retained. The
added training feature is `observation.images.cam_high_flow`. Wrist RGB cameras
remain available; only the global camera has optical flow.

Flow was estimated from the full original RGB pair, from frame `t` to
`min(t+8, episode_length-1)`. SAM foreground selection was applied to the resulting
dx/dy vectors. The stored flow video is the RGB visualization of those vectors,
not the original two-channel numerical vectors. Direction is hue; magnitude is
saturation with a fixed 0–20 pixel display scale; zero flow is white.

## Video locations

- Training flow: `videos/observation.images.cam_high_flow/chunk-000/file-NNN.mp4`
- Auxiliary binary videos: `mask_videos/cam_high/chunk-000/file-NNN.mp4`
- Original RGB: the existing `videos/observation.images.*` files

Each new video file contains 50 consecutive episodes. Video indices and episode
time offsets for the training feature are registered in `meta/episodes`. The
auxiliary videos are lossless and are deliberately absent from every feature,
video index and statistic in `meta`, so the standard LeRobot loader does not load
them. Their paths and time offsets are supplied separately in
`auxiliary_videos.json` at the dataset root.

## Validation and quality records

`merge_validation.json` records completed encoding, pixel and timestamp checks,
source preservation and official LeRobot loading checks. It appears only after
the merge has passed verification.

The original 357 automatically flagged episodes are retained, with their list in
`quality_review_list.csv`. Encoding does not repair or remove those episodes.

## State and action

Both `observation.state` and `action` are 14-dimensional joint-space vectors:
left arm (6 joints), left gripper (1), right arm (6 joints), right gripper (1).
They are not end-effector poses.

## Loading

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

repo_id = "tpy001/RoboTwin-Clean-RGB-Flow-Step8-LeRobot-v3"
root = HubApi().download_repo(repo_id, repo_type="dataset")
dataset = LeRobotDataset(repo_id, root=root)
```

There are 2,414 task descriptions. The dataset contains three original RGB camera
streams plus the global-camera flow visualization stream. No wrist optical flow
is included. The flow model is RAFT Large, and foreground masks use GroundingDINO
with SAM 2.1 propagation to select the robot and task objects.

The quality flags are automatic screening results, not a claim that every flagged
episode is defective. No episodes were removed. `completion_audit.json` retains
the original audit status; the supplied quality list omits server-local preview paths.