NilsCode's picture
Add dataset card and curation list
6c81a9a verified
|
Raw History Blame Contribute Delete
1.99 kB
metadata
license: cc-by-4.0
task_categories:
  - robotics
  - text-to-video
tags:
  - so101
  - 7dof
  - world-model
  - cosmos
  - video-generation

SO-101 cube pick-and-place clips for video world-model fine-tuning

These are 946 caption-paired video clips for fine-tuning video world models, such as NVIDIA Cosmos-Predict2.5 video2world with LoRA. The clips come from 473 teleoperated episodes: NilsCode/so101_cube_suite (171) and NilsCode/so101_rbw_100 (302).

Layout

videos/<source>__ep<NNNN>__<view>.mp4   # one clip per episode per camera, ~480p, h264 (crf 20)
metas/<source>__ep<NNNN>__<view>.txt    # caption for that clip
MANIFEST.json                            # source counts, clip count, target height
  • <source> is so101_cube_suite or so101_rbw_100, and <NNNN> is that dataset's episode index. You can join back to its joint states and actions.
  • <view> is front (static scene camera) or wrist (gripper camera). Each view is a separate clip, so a single LoRA learns to generate either view.
  • Each clip is the episode sliced from LeRobot's packed camera video and rescaled to 480 px height. Front clips from so101_rbw_100 are 16:9; those from so101_cube_suite are 4:3.

The caption is a view description plus the episode's task, e.g.:

Front camera view of a SO-101 robot arm on a tabletop. put the red cube in the bin. Robotic manipulation, static camera.

Robot and task

The robot is a custom 7-DOF SO-101 (6 joints + gripper), and the task is "put the {red, white, blue} cube in the bin". See the source datasets for joint order, cameras and curation lists. This dataset contains video and text only, no actions.

Associated reports

This data was used for a technical report on fine-tuning a video world model on a small custom-robot dataset. arXiv link to be added.

License

CC BY 4.0. Author: Nilesh Sutar.