JU9999/lewm_cube_single_expert_64px

LeWorldModel OGBench-Cube (cube_single_expert), 64 px

This is a redistribution of third-party data. We claim no ownership and make no endorsement claims; all credit to the original authors.

Upstream sources and licenses

• LeWorldModel (Maes, Le Lidec et al. 2026), https://huggingface.co/datasets/quentinll/lewm-cube (cube_single_expert.tar.zst), MIT, Copyright (c) 2026 Lucas Maes.
• Environment and scripted expert: OGBench (Park et al. 2024), https://github.com/seohongpark/ogbench, MIT, Copyright (c) 2024 OGBench Authors; collected with stable-worldmodel (https://github.com/galilai-group/stable-worldmodel, MIT, Copyright (c) 2026 GalilAI-group).

Source revision: quentinll/lewm-cube @ 02a19a67a0dc8c9d6215f89c19e0a597691e152a, file cube_single_expert.tar.zst. The full license texts are in LICENSE.

Modifications

• Frames resized 224 -> 64 px with antialiased bilinear filtering (torch F.interpolate, antialias=True), stored uint8 channels-first.
• Converted from the flat HDF5 (ep_len / ep_offset) to a TensorDict memmap of shape [episodes, longest episode]; shorter episodes are right-padded with copies of their last row and marked by mask = False (and truncated = True).
• Actions unchanged, except the NaN action the source stores on each episode's last row, which is stored as 0.0.
• No next/pixels / next/state.
• State = concat(qpos[21], qvel[20]); the per-episode target block pose (privileged_target_block_pos / yaw) is kept in data/episodes.npz; other columns (observation, proprio, prev_q, control, reward, render_time, ...) are dropped. Camera: the file's single front_pixels view.

Contents

TensorDict memmap (tensordict/) readable with tensordict.TensorDict.load_memmap, plus data/metadata.json (full provenance of the conversion) and data/episodes.npz. Converted with the BiLip-JEPA code base (src/bilip_jepa/envs/lewm/collect_data.py); load there with bilip_jepa.dataset_utils.resolve_dataset_ref({'kind': 'hf', 'repo_id': 'JU9999/lewm_cube_single_expert_64px', 'revision': 'v0'}).

• num_trajectories: 10000
• num_steps: 201
• image_shape: [3, 64, 64]
• raw_transition_count: 2010000
• padded_transition_count: 2010000
• dt: 0.05
• action_repeat: 1
• action_low: [-1.0, -1.0, -1.0, -1.0, -1.0]
• action_high: [1.0, 1.0, 1.0, 1.0, 1.0]
• action_min: [-1.0, -1.0, -1.0, -1.0, -0.6370766162872314]
• action_max: [1.0, 1.0, 1.0, 1.0, 0.9075843095779419]
• state_columns: ["qpos[0]", "qpos[1]", "qpos[2]", "qpos[3]", "qpos[4]", "qpos[5]", "qpos[6]", "qpos[7]", "qpos[8]", "qpos[9]", "qpos[10]", "qpos[11]", "qpos[12]", "qpos[13]", "qpos[14]", "qpos[15]", "qpos[16]", "qpos[17]", "qpos[18]", "qpos[19]", "qpos[20]", "qvel[0]", "qvel[1]", "qvel[2]", "qvel[3]", "qvel[4]", "qvel[5]", "qvel[6]", "qvel[7]", "qvel[8]", "qvel[9]", "qvel[10]", "qvel[11]", "qvel[12]", "qvel[13]", "qvel[14]", "qvel[15]", "qvel[16]", "qvel[17]", "qvel[18]", "qvel[19]"]
• trajectory_lengths_min_max_mean: [201, 201, 201.0]

Citation

Please cite the original works:

@article{maes_lelidec2026lewm,
  title={LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels},
  author={Maes, Lucas and Le Lidec, Quentin and Scieur, Damien and LeCun, Yann and Balestriero, Randall},
  journal={arXiv preprint arXiv:2603.19312},
  year={2026}
}

@article{park2024ogbench,
  title={OGBench: Benchmarking Offline Goal-Conditioned RL},
  author={Park, Seohong and Frans, Kevin and Eysenbach, Benjamin and Levine, Sergey},
  journal={arXiv preprint arXiv:2410.20092},
  year={2024}
}

File SHA-256 (as converted, before upload)

d115b675d7ec7108adcad6db4f366bc353e50c38ef691539adf10373361bbebc  data/episodes.npz
4f11c211cea695aca78a51e686d8863e32affe1901a20518d19f8a8b3f979858  data/metadata.json
5478eef6193c61a2f506efdaa3ac990661eab7e62bbf265d7372e5e4dcd47c5e  tensordict/action.memmap
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/done.memmap
97d8ae11c874dc6104ae7ced812d690352b70dff24b58493afd0dd3c8d3a5c98  tensordict/is_init.memmap
1e752b214eceb9aa1d5dc21c8ab022a49b2d8ce9fba024063937ce295b5a998f  tensordict/mask.memmap
1cf894eb5e152237cb22318e4eac845092a638b6e05b7391ba4426864a20217e  tensordict/meta.json
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/next/done.memmap
69c67f7d2576493f9523fd6ef4a919dd7861d92d6011ba64f510325e90df248d  tensordict/next/meta.json
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/next/terminated.memmap
a40a9bb511193707d2b016b21cbf7ec5d320765051470593b731022a6630c17a  tensordict/next/truncated.memmap
c7af264ae55f0e4ad7133bf2473c77ab2f72d4718425e24b0e94dcb6e8dc8b82  tensordict/pixels.memmap
c85fa75c3fc4fdb80332c6e1d425dd17b9ad874eb3af4c2fd31c8cecb046c106  tensordict/state/meta.json
405802fb4b65b6e4601b3cd8c9a9e0e66a806ec5c8df5e3bc6db9a91c49fe4bb  tensordict/state/observation.memmap
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/terminated.memmap
661f5dfd18e0302140926e9cbfd341857ca1af95b5a3c75e6dad5a503bf74569  tensordict/time.memmap
52e9745638630b0ca0b4e454d2965b692a9d74e9472d53a41e8127c304971bbd  tensordict/traj_ids.memmap
a40a9bb511193707d2b016b21cbf7ec5d320765051470593b731022a6630c17a  tensordict/truncated.memmap

Normalization statistics (stats/normalization.json)

Whole-dataset statistics: every split, real rows only, computed by scripts/compute_dataset_stats.py (module bilip_jepa.dataset_stats) of the ActionInjection_JEPA code. Training, the data loader, the planners and the evaluation build their scalers from this file.

• Data read: revision 137cedf862a23843f532d5259260779aae7a2493 (tag v0), layout rectangle; 10,000 episodes, 2,010,000 real rows (0 padding rows excluded), 2,000,000 action rows.
• Rows: states, pixels and references over rows t = 0 .. T_e - 1 of every episode e; actions over the transitions t = 0 .. T_e - 2 (the last row's action has no successor and is never read).
• Moments per coordinate j over those N rows, float64: mean_j = (1/N) sum_r x_rj, std_j = max(sqrt((1/N) sum_r (x_rj - mean_j)^2), 1e-6) (biased), min_j and max_j.
• Action percentiles p_low_j / p_high_j = the 0.1 % / 99.9 % percentiles of coordinate j over the action rows: sort x_(0) <= .. <= x_(N-1), h = (p/100)(N - 1), k = floor(h), q_p = x_(k) + (h - k)(x_(k+1) - x_(k)) (linear interpolation, numpy's default).
• Model units (z-score): x'_j = (x_j - mean_j) / std_j, inverse x_j = x'_j std_j + mean_j. Each scaler dict carries a units marker {format: bilip_jepa.affine_scaler, version: 1, kind, mode}.
• Planner bounds (raw; since stats file v2, 2026-09-27): low_j = max(p_low_j, env_low_j) and high_j = min(p_high_j, env_high_j) on coordinates where the environment enforces its declared limits, else low_j = p_low_j, high_j = p_high_j (v1 used min_j / max_j). In model units: [(low_j - mean_j) / std_j, (high_j - mean_j) / std_j]. Coverage: the fraction of action rows with every coordinate inside [low_j, high_j], and the box-area ratio prod_j (high_j - low_j) / (max_j - min_j).

Actions (5: dx_effector, dy_effector, dz_effector, dyaw_effector, dgripper):

• mean [0.01088, -0.003141, 0.002647, 0.0004239, 0.1593], std [0.2894, 0.3937, 0.6431, 0.3928, 0.2503]
• data min [-1, -1, -1, -1, -0.6371], data max [1, 1, 1, 1, 0.9076]
• 0.1 % percentile [-1, -1, -1, -1, -0.4176], 99.9 % percentile [1, 1, 1, 1, 0.6742]
• environment's declared limits [[-1, -1, -1, -1, -1], [1, 1, 1, 1, 1]], enforced by the environment: [false, false, false, false, false] (OGBench cube action_space Box(-1, 1); set_control does not clip the action)
• action rows outside the declared limits: 0 (0 %)
• planner bounds (raw) [[-1, -1, -1, -1, -0.4176], [1, 1, 1, 1, 0.6742]]; in model units [[-3.493, -2.532, -1.559, -2.547, -2.305], [3.418, 2.548, 1.551, 2.545, 2.057]]
• action rows inside the planner bounds: 1,996,000 of 2,000,000 (99.8 %); box-area ratio (planner box / data-range box) 0.7068
• no-actuation command (raw) [0, 0, 0, 0, 0] = [-0.03761, 0.007979, -0.004115, -0.001079, -0.6362] in model units, i.e. (raw - mean) / std (zero deltas: OGBench set_control target = current + a (gripper included))

State state/observation (41 raw columns: qpos[0], qpos[1], qpos[2], qpos[3], qpos[4], qpos[5], qpos[6], qpos[7], qpos[8], qpos[9], qpos[10], qpos[11], qpos[12], qpos[13], qpos[14], qpos[15], qpos[16], qpos[17], qpos[18], qpos[19], qpos[20], qvel[0], qvel[1], qvel[2], qvel[3], qvel[4], qvel[5], qvel[6], qvel[7], qvel[8], qvel[9], qvel[10], qvel[11], qvel[12], qvel[13], qvel[14], qvel[15], qvel[16], qvel[17], qvel[18], qvel[19]):

• angle columns [] (none) are mapped to (cos, sin) before the moments; 41 features: qpos[0], qpos[1], qpos[2], qpos[3], qpos[4], qpos[5], qpos[6], qpos[7], qpos[8], qpos[9], qpos[10], qpos[11], qpos[12], qpos[13], qpos[14], qpos[15], qpos[16], qpos[17], qpos[18], qpos[19], qpos[20], qvel[0], qvel[1], qvel[2], qvel[3], qvel[4], qvel[5], qvel[6], qvel[7], qvel[8], qvel[9], qvel[10], qvel[11], qvel[12], qvel[13], qvel[14], qvel[15], qvel[16], qvel[17], qvel[18], qvel[19]
• feature mean [-1.877, -1.551, 2.091, -2.105, -1.571, -0.3012, 0.2304, -9.559e-05, 0.229, -0.2247, 0.2304, -7.71e-05, 0.229, -0.2248, 0.4249, 0.0005164, 0.0691, 0.003743, -0.0002065, 0.0001557, 0.6002, 0.0008342, 0.03157, 0.02675, -0.04408, -2.232e-05, 0.0005436, 0.01258, 0.0106, 0.02094, -0.03679, 0.01258, 0.0106, 0.02095, -0.03686, -0.001826, 0.0005118, -0.0001095, -3.667e-05, -5.527e-05, 0.0006062], std [0.3647, 0.2135, 0.2183, 0.2596, 0.0001955, 1.753, 0.2003, 0.00361, 0.1989, 0.1964, 0.2003, 0.002367, 0.1988, 0.1965, 0.07239, 0.1694, 0.08248, 0.7067, 0.01672, 0.01803, 0.3737, 0.3076, 0.3759, 0.3853, 0.4739, 0.00431, 0.5119, 0.4392, 0.08632, 0.467, 0.5275, 0.4392, 0.08548, 0.4669, 0.5271, 0.05384, 0.07018, 0.1526, 0.1786, 0.1789, 0.2596], min [-2.822, -2.287, 1.048, -2.528, -1.578, -6.283, 0, -0.6289, -0.2627, -0.7582, 0, -0.6273, -0.004045, -0.7627, 0.2444, -0.3556, 0.013, -1, -0.6943, -0.7885, -1, -1.731, -1.835, -1.755, -2.545, -0.1815, -2.211, -2.542, -12.9, -10.74, -10.03, -2.542, -6.693, -5.324, -10.05, -0.7984, -0.698, -2.218, -19.9, -17.89, -31.38], max [-0.9577, -1.03, 2.689, -1.288, -1.554, 6.283, 0.7822, 0.008396, 0.7832, 0.8966, 0.7821, 0.008392, 0.781, 0.7934, 0.6108, 0.355, 0.348, 1, 0.8441, 0.7991, 1, 1.7, 1.558, 1.919, 1.878, 0.1601, 2.037, 4.293, 4.371, 4.206, 26.29, 4.29, 4.77, 3.985, 17.42, 0.713, 0.5966, 0.7324, 16.77, 15.9, 24.54]
• raw column min [-2.822, -2.287, 1.048, -2.528, -1.578, -6.283, 0, -0.6289, -0.2627, -0.7582, 0, -0.6273, -0.004045, -0.7627, 0.2444, -0.3556, 0.013, -1, -0.6943, -0.7885, -1, -1.731, -1.835, -1.755, -2.545, -0.1815, -2.211, -2.542, -12.9, -10.74, -10.03, -2.542, -6.693, -5.324, -10.05, -0.7984, -0.698, -2.218, -19.9, -17.89, -31.38], max [-0.9577, -1.03, 2.689, -1.288, -1.554, 6.283, 0.7822, 0.008396, 0.7832, 0.8966, 0.7821, 0.008392, 0.781, 0.7934, 0.6108, 0.355, 0.348, 1, 0.8441, 0.7991, 1, 1.7, 1.558, 1.919, 1.878, 0.1601, 2.037, 4.293, 4.371, 4.206, 26.29, 4.29, 4.77, 3.985, 17.42, 0.713, 0.5966, 0.7324, 16.77, 15.9, 24.54]

Pixels: per channel over all real frames of x / 255: mean [0.1875, 0.2186, 0.2957], std [0.1121, 0.1084, 0.1171].