ChongkaiGao/ManiLadder_preprocessed_dataset

ManiLadder Preprocessed Dataset

Preprocessed robotic manipulation demonstrations, normalization statistics, and camera videos from ManiLadder.

This repository contains the complete preprocessed/ snapshot from Linslab/ManiLadder_Dataset, revision 8526d8017bdc87a7ec97cc715928792cfd4b5107, migrated on 2026-10-09. The migrated snapshot contains 99 task directories and 12,473 data files, with a total logical file size of approximately 3.69 TB.

Layout

Task directories are at the repository root:

<task>/
  norm_stats.json
  traj_0.h5
  traj_1.h5
  ...
  traj_0_front_camera.mp4
  ...

Camera names and trajectory counts vary by task. The source directory contents are preserved, including any existing additional statistics files and nested directories. The outer preprocessed/ prefix is omitted in this repository.

All 12,360 binary files were checked against the original SHA-256 hashes and sizes; the 113 other files were checked against their original Git blob hashes. HDF5 and video downloads were sampled for every task.

Download a task

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ChongkaiGao/ManiLadder_preprocessed_dataset",
    repo_type="dataset",
    allow_patterns="BuildTent-v1/*",
    local_dir="./dataset",
)

Replace BuildTent-v1 with the task directory you need. For an individual file, use a path such as BuildTent-v1/traj_0.h5.

Related artifacts

• Raw scripted demonstrations
• Trained checkpoints

The original dataset's MIT license is retained.

Additional dataset: DualArmSandShovel-v1

The revised SandShovel dataset is published under DualArmSandShovel-v1/. It contains the exact 20 demonstrations used to train the 40,000-step checkpoint: 20 HDF5 trajectories, 100 camera videos, and norm_stats.json with the floored normalization statistics used during training. The experiment is sandshovel_v2/run_20261005_a.

• Raw demonstrations and metadata
• 40,000-step checkpoint

Including SandShovel, the repository contains 100 task directories and 12,594 data files. All 243 newly published SandShovel artifact files across the three locations were verified against their original sizes and hashes.