NFWM World-Model Datasets

Current training and evaluation artifacts for NFWM, LeWM, and SMWM comparisons. The four standard environments use episode-disjoint SMWM-style splits; Teleport uses V5 exact-goal training and Fixed-Y evaluation; LIBERO uses the regenerated white-mug training set and fixed pi0.5 noisy candidate pool. Stochastic Boxing uses the unchanged 50% P2-replacement dataset and the shared model implementation.

Pretrained Models

The checkpoint index provides 21 pretrained world models: NFWM, LeWM, and SMWM for six robotics environments and Stochastic Boxing. Use the reproduction commands to download a task and evaluate it without local cluster paths. Epochs are 10 for the four standard tasks, 150 for Teleport V5, and 30 for LIBERO. Boxing also includes predictor reward/termination heads, calibration, and final 20K imagination-RL policies (no BC initialization weights). See the Boxing recipe.

Dataset Index

Environment	Training	Held-out evaluation	Protocol
PushT	16,816 episodes	1,869 episodes; 50 fixed planning slices	90/10 episode split, seed 3072
Cube	9,000 episodes	1,000 episodes; 50 fixed planning slices	Same split rule
Reacher	9,000 episodes	1,000 episodes; 50 fixed planning slices	Same split rule
TwoRooms	9,000 episodes	1,000 episodes; 50 fixed planning slices	Same split rule
Teleport TwoRooms V5	20,000 episodes	25 upper/lower pairs, 50 episodes	Fixed target y=112; final 25-step windows
LIBERO white mug	18,685 regenerated episodes	50 successful scenes, 50 candidates per scene	Fixed pi0.5+noise action selection
Stochastic Boxing	660 episodes	70 episodes	See the original Boxing card

Standard Environments

Each directory contains train.h5, eval.h5, split_manifest.json, and eval_manifest.json. Source episodes come from the original LeWM datasets; the split reproduces the default episode splitter in sensorimotor-world-model:

permutation = np.random.default_rng(3072).permutation(num_episodes)
cut = int(round(0.9 * num_episodes))
train_episodes = np.sort(permutation[:cut])
eval_episodes = np.sort(permutation[cut:])

Split manifests retain the original episode IDs. Within each physical HDF5, episode indices and offsets refer to the local split. The evaluation manifest pins the exact 50 rows, episode IDs, start steps, and original source episode IDs used across models and checkpoints. Selection seed is 42. This is not the early 21-episode PushT validation file, nor evaluation on the complete training file.

The stored RGB images are 224x224. Model-specific resizing, normalization, augmentation, input noise, and rollout randomness are applied by the model code, not baked into this release. Action alignment and all state columns are preserved. The standard planning manifests specify CEM 300/30/30, goal offset 25 raw steps, action blocks of five raw steps, planning horizon five blocks, and a 50-raw-step execution budget. An identical manifest does not by itself fix planner seeds or model-side noise settings; record those separately when reproducing results.

These held-out episodes are also used for checkpoint screening in the current experiments. They should not be represented as a second untouched test set after model selection. They reproduce the splitter, not necessarily every upstream paper's full training/evaluation recipe or any newly downloaded upstream revision.

Teleport V5

• Training HDF5: 9,000 exact-goal expert episodes, 9,000 opposite-branch action replays, and 2,000 collision episodes.
• Fixed-Y evaluation HDF5: 25 paired initial scenes, evaluated under both upper and lower teleport branches. There is no middle teleport branch. Target y is exactly 112; target x varies in [180, 202].
• Evaluation manifest: all 50 final windows, each spanning 25 raw transitions and containing teleportation.
• Evaluation uses an exactly rendered canonical position-target goal and a one-shot open-loop 25-step plan, not replanning. The standard environment speed is 5 and frame skip is 5. The generation record retains reset seeds and pairing.

Counterfactual replay uses the other branch's action suffix under the actual branch dynamics. Such trajectories are not necessarily all failures. The exact-goal training generation record includes its historical variable-Y evaluation source; that reference is provenance only. The active test split in this release is fixed_y/eval.h5 and its manifest, not that historical source.

The supplementary datasets/teleport/training_validation.h5 restores the historical 50-episode variable-target validation view from the archived release for SMWM training compatibility. It is not the Fixed-Y planning test set.

LIBERO White Mug

Task: put the white mug on the plate, using the custom two-object scene. This release replaces the former LIBERO-90 mixed-task training artifact.

Training

training/train.h5 contains 18,685 pi0.5 episodes: 13,080 successes and 5,605 failures. It stores 128x128 RGB pixels (agentview) and pixels_wrist (robot0_eye_in_hand), seven-dimensional actions, proprioception, end-effector pose, gripper state, joint positions/velocities, scene perturbations, episode seeds, and initialization records. proprio has eight dimensions; the additional named robot-state columns are retained too.

Actions are (dx, dy, dz, daxis_angle_x, daxis_angle_y, daxis_angle_z, gripper). Observations are pre-action: row t stores observation o_t and action a_t. No language input is required by the world models. Single-camera comparisons read only pixels and action; optional wrist/proprio fields do not silently change the model's input. This is the later regenerated dataset, not a claim that every historical checkpoint used these exact regenerated episodes.

Fixed Action-Selection Test Pool

The complete pool card and files are retained byte-for-byte:

• 50 successful reference episodes; 50 fixed pi0.5-generated noisy candidates per scene, totaling 2,500 candidates.
• Mean native candidate success is 26.04%. Every scene has at least one successful candidate, giving a 100% oracle ceiling.
• No ground-truth action sequence is inserted. All candidates have nonzero action noise. This is not the earlier GT-action-plus-noise benchmark.
• Success uses the native LIBERO task predicate during full-task execution, not a custom pixel-distance or intermediate-position threshold.
• Cached outcomes correspond to the tail-padded-to-five action protocol.
• The candidate pool remains fixed across models. NFWM input/rollout seeds may change without regenerating the action pool.

Training and test episode seeds are disjoint. They share the task, scene perturbation family, and expert family, but training is quota-selected 70/30 success/failure while test references are success-conditioned. The candidate pool is also difficulty-calibrated. This is a conditional ranking benchmark, not an unbiased estimate of pi0.5's overall policy success. The fixed test reference file is single-camera; wrist/proprio training fields do not imply that extra observations are present in the test candidate pool.

Stochastic Boxing

The Boxing dataset card, train/test HDF5s, manifest, and checksums are unchanged by this refresh. Please refer to that card for ROM/environment requirements and collection. The Boxing checkpoint release adds final world models, predictor reward/termination heads and imagination-RL policies. No Atari ROM, PPO collector policy or BC initialization is shipped.

Download and Read

These are physical HDF5 datasets, not Parquet tables; use h5py, not the generic Hugging Face tabular Dataset Viewer. Download only the environment you need:

from huggingface_hub import snapshot_download
import h5py
import hdf5plugin  # Required for original Teleport compression filters.

root = snapshot_download(
    repo_id="PeterLiZijia/nfwm", repo_type="dataset",
    allow_patterns=["datasets/pusht/*", "SHA256SUMS"],
    local_dir="nfwm-data",
)
with h5py.File(f"{root}/datasets/pusht/train.h5", "r") as data:
    print(data["pixels"].shape, data["action"].shape)

Metadata preserves historical absolute cluster paths for provenance. Resolve the corresponding files using the paths in this card after downloading; those old absolute paths are not external HDF5 data dependencies. Simulator-based replay additionally needs the corresponding environment and reset adapters.

Integrity and Versions

SHA256SUMS records uploaded-file hashes. Current audit records split, pairing, and candidate-pool checks. Logical integrity records per-column hashes for losslessly repacked files. Standard train/eval files and regenerated LIBERO training use gzip compression for transfer; every data value, dtype, shape, ordering, and attribute is preserved and verified. HDF5 container bytes therefore differ from the original uncompressed/local containers. There is no resizing, re-splitting, augmentation, or action modification during repacking.

Obsolete unsplit files, LIBERO-90 training, and variable-Y Teleport tests are no longer the main-branch defaults. The tag before-current-datasets-20260919 preserves the previous repository revision. Existing Boxing and checkpoint artifacts are not replaced; the Boxing model/policy bundle is an addition. Upstream datasets and simulators retain their original licenses and attribution; this collection does not relicense them.