STDB-CC-datasets / README.md
lyang116's picture
Add files using upload-large-folder tool
c7f8864 verified
|
Raw
History Blame Contribute Delete
1.99 kB
metadata
license: apache-2.0
task_categories:
  - robotics
  - imitation-learning
pretty_name: STDB-CC Datasets

STDB-CC Datasets

STDB-CC is a RoboTwin-based robot imitation learning dataset for studying visual distribution shift and task/environment factorization.

The dataset uses one fixed split:

scene0 + scene3 -> train
scene2          -> validation
scene1          -> test

Split token:

scene03_train_scene2_val_scene1_test

Directory Layout

raw/robotwin_stdb_cc/
  <task_name>/
    stdb_cc_scene0/
    stdb_cc_scene1/
    stdb_cc_scene2/
    stdb_cc_scene3/
    manifests/scene03_train_scene2_val_scene1_test/

processed/diffusion_policy/
  scene03_train_scene2_val_scene1_test/
    <task_name>-stdb_cc_task-50.zarr/

metadata/
  dataset_summary.json
  verify_reports/

Contents

  • raw/robotwin_stdb_cc/: raw RoboTwin STDB-CC data with 4 visual scenes per task.
  • processed/diffusion_policy/: Diffusion Policy zarr datasets built from the train split only.
  • metadata/verify_reports/: local verification reports for the raw task data.
  • metadata/dataset_summary.json: machine-readable task list, split information, and artifact counts.

Current Release

  • Tasks: 50
  • Episodes per task per scene: 50
  • Raw episodes per task: 200
  • DP train episodes per task: 100
  • Scenes: stdb_cc_scene0, stdb_cc_scene1, stdb_cc_scene2, stdb_cc_scene3

The raw task manifests are stored under each task:

raw/robotwin_stdb_cc/<task_name>/manifests/scene03_train_scene2_val_scene1_test/

The DP zarr artifacts embed their train manifest as:

processed/diffusion_policy/scene03_train_scene2_val_scene1_test/<task_name>-stdb_cc_task-50.zarr/stdb_cc_manifest.jsonl

Notes

This repository currently contains the raw RoboTwin STDB-CC data and the Diffusion Policy conversion. Future conversions for other policy families can be added under processed/<format_or_model>/ without changing the raw data layout.