jzndd/RoboTwin_adjust_bottle_For_ACWM_Training

RoboTwin adjust_bottle for ACWM Training

中文说明 · Training code · WoVR+ project

A public downstream dataset for training and testing action-conditioned world models on the RoboTwin adjust_bottle task. It contains 500 policy-rollout episodes, 121,263 synchronized timesteps, three RGB cameras, joint actions, end-effector poses, and camera calibration. The HDF5 payload totals 5,676,701,632 bytes (5.68 GB / 5.29 GiB). There are no pretrained model weights or latent caches in this dataset.

Task instruction: “Pick up the long thin bottle upright from the table.” Observations are simulator camera renderings; these files are not world-model-generated videos. The policy identifier stored in the files is pi05. Both successful and unsuccessful rollouts are retained; this is not a success-only expert-demonstration collection.

Rollout generation

The directory name determines the generation group. Following the dataset author's WoVR+ settings, the release uses chunkwise-noise, level=0.2, steps=5. Chunk-wise noise couples exploration across neighboring actions in a chunk. The unsuffixed group is the ODE-rollout baseline; the four numbered groups are SDE rollouts. These labels describe action-policy sampling, not VAE encoding or world-model initialization.

Directory	Rollout	Episodes	Timesteps	Successes	Frame range
adjust_bottle	ODE	100	21,100	79	110–400
adjust_bottle_1	SDE repetition 1	100	26,166	57	119–400
adjust_bottle_2	SDE repetition 2	100	25,226	62	95–400
adjust_bottle_3	SDE repetition 3	100	24,918	62	101–400
adjust_bottle_4	SDE repetition 4	100	23,853	68	98–400

All 400 SDE files explicitly record noise_level=0.2, sample_num_steps=5, and sde_chunkwise_noise=true. The ODE files do not repeat these fields: the shared settings for that group are author-supplied metadata, and should not be interpreted as evidence that the ODE baseline injected SDE noise. See generation.json. Some source SDE files retain a legacy repeat_index=0; use the directory group, not that field, to identify repetitions.

Download

pip install huggingface_hub h5py pillow numpy
hf download jzndd/RoboTwin_adjust_bottle_For_ACWM_Training \
  --repo-type dataset --local-dir /your/data/adjust_bottle

Or use snapshot_download(..., repo_type="dataset", local_dir=...) from huggingface_hub. Data should live outside the training Git checkout. Pin a dataset commit with --revision for reproducible experiments.

Layout and splits

adjust_bottle/episode0.hdf5 ... episode99.hdf5
adjust_bottle_1/ ... adjust_bottle_4/
manifest.json           # per-file SHA256, array digest, size, outcome and split
generation.json         # rollout settings and verification scope
schema-example.json     # concrete HDF5 dataset shapes and dtypes
splits/train.txt        # 400 relative paths
splits/validation.txt   # 100 relative paths

The same episode number across the five groups shares a scene seed. Split by scene, not by randomly distributing files: episode IDs 0–79 in every group are training, and 80–99 in every group are validation. This gives 80 training scenes and 20 held-out scenes, each with five rollouts. A random per-file split would leak scene layouts between sets. The supplied training YAMLs implement this split. These are downstream development splits, not an independent standardized leaderboard or evidence of unseen-task generalization.

HDF5 schema

N is the episode's timestep count. RGB values are individually JPEG-encoded byte strings, not dense N×H×W×3 arrays.

Key	Shape / meaning
action, joint_action/vector	N×14, left arm 6 + left gripper + right arm 6 + right gripper
joint_action/left_arm, right_arm	N×6 joint targets
joint_action/left_gripper, right_gripper	N gripper values
endpose/left_endpose, right_endpose	N×7 stored end-effector poses; use the supplied adapter's convention
observation/{head_camera,left_camera,right_camera}/rgb	N JPEG byte strings
.../intrinsic_cv	N×3×3 camera intrinsics
.../extrinsic_cv	N×3×4 camera extrinsics
.../cam2world_gl	N×4×4 OpenGL camera-to-world matrices
pointcloud	N×0 placeholder; no point-cloud observations provided

import io
import h5py
import numpy as np
from PIL import Image

with h5py.File("/your/data/adjust_bottle/adjust_bottle/episode0.hdf5", "r") as f:
    image = Image.open(io.BytesIO(bytes(f["observation/head_camera/rgb"][0]))).convert("RGB")
    actions = np.asarray(f["action"], dtype=np.float32)
    print(image.size, actions.shape, bool(f.attrs["success"]))

Use the release's robot embodiment assets and forward-kinematics adapter for action projection; joint vectors alone do not specify camera-space geometry. Robot meshes/URDFs are downloaded separately from the upstream RoboTwin asset repository.

Integrity and preprocessing

Every HDF5 dataset array was compared byte-for-byte with its source during release preparation, including JPEG payloads, actions, poses, and calibration arrays. The only HDF5 metadata change is normalization of SDE reference_path attributes to release-relative paths. Public summaries were regenerated to remove internal machine paths. Sources were not modified.

manifest.json provides published-file SHA256 values; the training repository includes scripts/verify_dataset.py to stream-check all 500 files. Latent caches are intentionally excluded: generate them using the training code and the desired sampling/augmentation settings. A cached augmentation version fixes a sampled window transformation, not the diffusion noise or diffusion timestep used during training.

Intended use and limitations

• Train and compare action-conditioned video/world models; inspect action representations and evaluate held-out prediction loss.
• One simulated task, one embodiment, one policy family, and only 100 scene seeds. Real-robot transfer and broad task generalization are not established by this release.
• Success labels describe the recorded rollouts, not the performance of a subsequently trained world model.
• Frames are temporally correlated. Report the scene split, rollout groups, context/horizon, view ordering, augmentations, and random seeds when comparing models.
• No real-person recordings are included; the visible scene consists of the simulated robot, objects, and environment.

License and attribution

This dataset is released by jzndd / youmo445 under CC BY 4.0. Attribute this dataset and link to its version when redistributing or publishing results; indicate modifications. Upstream RoboTwin code/assets and separately downloaded model weights retain their own licenses. The accompanying training code is Apache-2.0.

@misc{youmo4452026robotwinadjustbottle,
  author = {youmo445},
  title = {RoboTwin adjust\_bottle for Action-Conditioned World Model Training},
  year = {2026},
  url = {https://huggingface.co/datasets/jzndd/RoboTwin_adjust_bottle_For_ACWM_Training}
}

Acknowledgements: RoboTwin, WoVR+, and the upstream policy and simulation ecosystem. This dataset card does not assign a paper identifier to unpublished work.