README.md

metadata

license: mit
task_categories:
  - robotics
tags:
  - point-cloud
  - world-model
  - canonical-point-cloud
  - rigid-transform
pretty_name: 3DWM kinder training datasets

3DWM training datasets

Demonstration data converted to the point-cloud dataset format used by the 3DWM world model, i.e. the layout its maniskill loader expects:

<dataset>/<split>/<scenario>/demo_N/
    phases_dict.pkl      {'cannonical_points': {geom: (P, 3)}}   canonical local-frame points
    0.h5 .. (T-1).h5     transforms/<geom> (4, 4), action (7,), names, tcp_pose (7,)

Points are stored once per geom in its local frame, with a 4x4 local-to-world transform per timestep, which gives point-to-point correspondence across time, clean part segmentation and complete surfaces. Velocity features are computed by differencing corresponding points, so these are not sensor point clouds and a raw depth capture cannot be substituted.

Contents

• StackCube-v1-demos/ — ManiSkill 3 StackCube, unpacked per-demo directories.
• Franka-pickplace-1000demos-v2.tar.gz — 74 MB, md5 23decfb43157934b6a4cfa1871dac1af. kinder MuJoCo FrankaPickPlace3D-o1: 1000 train + 10 test demos, 111,181 files, 1.3 GB unpacked. Shipped as a tarball because ~110k files of ~12 KB each is a poor fit for per-file hosting.

hf download Flashkernel/3dwm-kinder-data Franka-pickplace-1000demos-v2.tar.gz \
    --repo-type dataset --local-dir .
tar xzf Franka-pickplace-1000demos-v2.tar.gz -C data/

Sanity check after extracting — part names, order and counts must match what the training config lists as env_keys:

import pickle
d = pickle.load(open('data/Franka-pickplace-1000demos-v2/train/pickplace/demo_0/phases_dict.pkl','rb'))
pts = d['cannonical_points']
print(list(pts.keys()))
print([v.shape[0] for v in pts.values()])   # [300, 38,38,38,38, 37,37,37,37]

Franka pickplace provenance

Converted from the raw demos with data_generation/kindergarden/convert_sweep_to_3dwm.py --pads 300 --cube 300, which selects the gripper pads and the cube and sets their point budgets. Actions are normalized with pos_scale=0.1 and rot_scale=-0.1 (note the sign) and clipped to [-1, 1]; real end-effector motion is ~14 mm per step.

Earlier stages of the chain — the raw kindergarden demo pickles and the packed canonical HDF5 — are under franka_pickplace/ in Flashkernel/Kinder-worldmodel. The world model trained on this data is at Flashkernel/3dwm-franka-pickplace-mppi, whose card documents which kindergarden commit the environment must be pinned to (it matters, and getting it wrong fails silently).