yixuan1999/interactive-world-sim-expert-data

Interactive World Simulator — Expert Demonstrations

Original expert-demo HDF5 trajectories recovered from the project's data/real_aloha archive. This dataset complements the world-model training dataset.

Upload in progress. manifest.csv lists the intended complete release; trajectory files appear as each batch finishes.

Layout and source versions

Files follow <task>/<split>/episode_<index>.hdf5, matching the world-model dataset's directory layout. The HDF5 files are uploaded byte-for-byte without conversion, resizing, filtering, or metadata changes. Episode numbering is local to each split.

Task	Source directory	Train episodes	Validation episodes
bimanual_box	bimanual_box_demo_1223	100	3
bimanual_rope	bimanual_rope_demo_0108	100	—
bimanual_sweep	bimanual_sweep_demo_0114	100	—
single_chain_in_box	single_chain_in_box_demo_0114	100	—
single_grasp	single_grasp_demo_0114	100	—

The four flat source directories are placed under train/; no new validation split is created. manifest.csv records each uploaded path, original relative source path, and byte size. There are 503 trajectories in total.

Contents

Sampled episodes contain action, joint_action, obs/joint_pos, obs/full_joint_pos, obs/ee_pos, obs/world_t_robot_base, timestamp, and two camera views under obs/images/, including camera intrinsics and extrinsics. Sampled color images are uint8 arrays shaped (T, 480, 640, 3). Action dimensions differ by task; consult the actual file rather than assuming a shared action representation.

import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yixuan1999/interactive-world-sim-expert-data",
    repo_type="dataset",
    filename="single_grasp/train/episode_0.hdf5",
)
with h5py.File(path, "r") as episode:
    actions = episode["action"][:]
    images = episode["obs/images/camera_0_color"][:]

Provenance notes

Sample frames show physical robot recordings. However, the sampled original demo files carry a root attribute sim=True; this inconsistent attribute is preserved as found. These archived source versions have not been cross-checked against final paper training-run records.

This release contains the five identified expert-demo task datasets. PushT expert demonstrations were not separately identified and are not included. World-model-generated demonstrations, mixture datasets, converted LeRobot variants, redundant standalone videos, and DP/ACT training configurations are not included.