snakerex/episodes-offset25-400

PushT, TwoRoom, Reacher and Cube: offset 25

400 trajectory segments, 100 per environment. Each segment contains 26 original observations and exactly 25 actions. The goal is the last observation, exactly 25 environment steps after the first observation. Segments are extracted from existing trajectories; they are not new policy rollouts. Native image resolution and state/action values are preserved without normalization.

Sampling uses NumPy default_rng(42): 100 distinct source episodes per environment, then a uniformly sampled valid start within each episode. PushT, TwoRoom and Cube use the full source episode pool; Reacher uses the first 300 source episodes, streamed from the official archive. Its exported HDF5 chunks are checked to lie entirely inside the populated source prefix. Sampling is not conditioned on task success. A terminal flag at the final observation is preserved; no segment crosses a source episode boundary.

Files

• episodes/ENV/episode_NNN.npz: one segment per file, loadable without pickle.
• manifest.json: source inventory, shapes, dtypes, episode coordinates and hashes.
• manifest.jsonl: one row per segment, with paths relative to the archive root.
• build_episodes.py: extraction, validation and packaging script.
• SHA256SUMS: SHA-256 checksums for the extracted payload.

All transition-indexed source fields are retained. pixels has 26 frames and action has 25 rows: action[t] takes observation[t] to observation[t+1]. Other original fields have 26 rows. Original step_idx and ep_idx/episode_idx are preserved. segment_step_idx starts at zero. Scalar metadata records the original episode, start and goal. String fields are stored as NumPy strings rather than object arrays. Unavailable diagnostic success/score/render_time values are preserved as source NaNs and counted in each manifest entry; observation, action and state values are required to be finite.

import numpy as np
with np.load("episodes/pusht/episode_000.npz", allow_pickle=False) as episode:
    start_image = episode["pixels"][0]
    goal_image = episode["pixels"][-1]
    actions = episode["action"]
    assert len(actions) == 25
    assert len(episode["pixels"]) == 26

Environment state fields include PushT state, TwoRoom pos_agent and pos_target, and Cube/Reacher qpos and qvel when available in the source. Consult each environment's reset/set_state API when restoring simulator state.

Source datasets are the LeWorldModel datasets, with provenance in manifest.json. All four upstream repositories declare the MIT license. Original project: https://huggingface.co/collections/quentinll/lewm

Validation checks every exported field against its source slice, NPZ readback, episode boundaries, action/observation counts, finite observation/action/state values, and hashes. Diagnostic NaNs are compared with NaN-aware equality.

Reproduction requires Python, NumPy, h5py and hdf5plugin. The latter registers the Blosc compression filter needed to read the original source files.