Datasets:

jeffwang0303
/

test

Modalities:

Video

Size:

1K - 10K

Libraries:

Datasets

Dataset card Data Studio Files Files and versions

xet

Community

Episodes Preview realman_65 Visualizer Open in visualizer

1k episodes · 15 fps · 3 cameras · 640×480 h264

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

RoboRender augmentation of the Realman CMP-cover task — 1000 episodes

Source: 3 real teleoperated episodes, task 1177 外骨骼02/CMP上盖摆料/20260819任务二 (runs data4 / data30 / data53). Robot: Realman 65 dual-arm. Cameras: camera_main (static third-person), camera_left and camera_right (wrist).

Each episode is one RoboRender re-rendering of a source run with its own scene prompt (background, table surface, lighting).

Model: Sim2RealGen = Wan2.1-Fun-V1.1-1.3B-Control + LoRA 2.8 (droidnew_continue, step-5294).
Conditioning: VDA depth + SAM3 robot-only mask, 3 views jointly, 81-frame clips chained on the previous clip's last frame.
Sampling: TeaCache 0.10, 50 steps, cfg 5.

Robot motion, actions and states are the source run's own.

LeRobot v2.1 layout (meta/info.json, meta/episodes.jsonl, data/chunk-000/…parquet, videos/chunk-000/observation.images.<cam>/…mp4)
15 fps: every 2nd source frame; actions and states are subsampled the same way.
Videos are 640x480. They were generated at 416x240, the model's native size, and resized back.
Episode i comes from source run [data4, data30, data53][i % 3], so lengths are 713 / 771 / 343 frames.
gen/episode_XXXXXX.json and meta/generation.jsonl hold the prompt, seed and source run of every episode.
Known limitation: in the longest episodes (data30, 10 clips) the background softens slightly after clip ~6 (chained-reference drift). The robot and the pallet stay intact.

Downloads last month: 60