JU9999/lewm_reacher_64px

LeWorldModel DMC Reacher (random actions), 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-reacher (reacher.tar.zst), MIT, Copyright (c) 2026 Lucas Maes.
• Environment: DeepMind Control Suite reacher (Tassa et al. 2018) on MuJoCo, Apache-2.0; collected with stable-worldmodel (MIT, Copyright (c) 2026 GalilAI-group), uniform random actions, action repeat 2 (dt 0.04 s), target disc NOT rendered (alpha 0).

Source revision: quentinll/lewm-reacher @ e70a080d0d04c6072123c9ebd343acf7fff28dbf, file reacher.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 = [qpos(2), qvel(2), target_pos(2)]; other columns (observation, finger_pos, reward, score, success, render_time, id, ...) dropped. Target visibility: hidden (as in the source frames).

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_reacher_64px', 'revision': 'v0'}).

• num_trajectories: 10000
• num_steps: 201
• image_shape: [3, 64, 64]
• raw_transition_count: 2010000
• padded_transition_count: 2010000
• dt: 0.04
• action_repeat: 2
• action_low: [-1.0, -1.0]
• action_high: [1.0, 1.0]
• action_min: [-0.9999997615814209, -0.9999998211860657]
• action_max: [0.9999979138374329, 0.9999991059303284]
• state_columns: ["qpos[0]", "qpos[1]", "qvel[0]", "qvel[1]", "target_pos[0]", "target_pos[1]"]
• 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{tassa2018deepmind,
  title={DeepMind Control Suite},
  author={Tassa, Yuval and Doron, Yotam and Muldal, Alistair and Erez, Tom and Li, Yazhe and de Las Casas, Diego and Budden, David and Abdolmaleki, Abbas and Merel, Josh and Lefrancq, Andrew and Lillicrap, Timothy and Riedmiller, Martin},
  journal={arXiv preprint arXiv:1801.00690},
  year={2018}
}

File SHA-256 (as converted, before upload)

8cae7fa7290d742c83db0e2b9c786cd251c894efd8d52b8a92d31ee8a941d23e  data/episodes.npz
b79f77a1af3ce42fda95557871604785e58a47eedd8f35b416cda43d66b742c7  data/metadata.json
912e82536d67a04734c66966e5f042389ea826ea0e349b37b99dbaa88c0330c1  tensordict/action.memmap
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/done.memmap
97d8ae11c874dc6104ae7ced812d690352b70dff24b58493afd0dd3c8d3a5c98  tensordict/is_init.memmap
1e752b214eceb9aa1d5dc21c8ab022a49b2d8ce9fba024063937ce295b5a998f  tensordict/mask.memmap
f5f50fe086b7b5a4041c0479dd1f20c50628895c1d4331d6ba664740c58cb07e  tensordict/meta.json
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/next/done.memmap
69c67f7d2576493f9523fd6ef4a919dd7861d92d6011ba64f510325e90df248d  tensordict/next/meta.json
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/next/terminated.memmap
a40a9bb511193707d2b016b21cbf7ec5d320765051470593b731022a6630c17a  tensordict/next/truncated.memmap
c285b9e444066166c7e6ab5fe5f22469cb8389f067b652cd12ee155781428b05  tensordict/pixels.memmap
e45b4b8b5632a6cda21542c2d603f31433f3bb2893c15b19eada1386323d90ca  tensordict/state/meta.json
e6bc16113630c514fd04a35bdff0d2fa1189c4ee10d463b41b43c1dd630d8719  tensordict/state/observation.memmap
b936685a9a3b724848f0ce19ca9bb014db05b737714672916c841a03c63ff001  tensordict/terminated.memmap
c4a32a31d7dbcdc582bf9048fcefadd8a945e46448d6b5e7f299bc368ac21b9e  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 eb1ffaee249f22bbd468dd346a2719f847898382 (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 (2: tau_shoulder, tau_wrist):

• mean [3.573e-05, -0.0005056], std [0.5774, 0.5772]
• data min [-1, -1], data max [1, 1]
• 0.1 % percentile [-0.998, -0.9979], 99.9 % percentile [0.998, 0.9981]
• environment's declared limits [[-1, -1], [1, 1]], enforced by the environment: [true, true] (dm_control reacher.xml ctrllimited ctrlrange -1 1)
• action rows outside the declared limits: 0 (0 %)
• planner bounds (raw) [[-0.998, -0.9979], [0.998, 0.9981]]; in model units [[-1.728, -1.728], [1.728, 1.73]]
• action rows inside the planner bounds: 1,992,010 of 2,000,000 (99.6 %); box-area ratio (planner box / data-range box) 0.996
• no-actuation command (raw) [0, 0] = [-6.188e-05, 0.000876] in model units, i.e. (raw - mean) / std (zero joint torques)

State state/observation (6 raw columns: qpos[0], qpos[1], qvel[0], qvel[1], target_pos[0], target_pos[1]):

• angle columns [0, 1] (qpos[0], qpos[1]) are mapped to (cos, sin) before the moments; 8 features: cos(qpos[0]), sin(qpos[0]), cos(qpos[1]), sin(qpos[1]), qvel[0], qvel[1], target_pos[0], target_pos[1]
• feature mean [-0.01298, -0.01008, 0.135, -0.006901, 0.0001134, -0.0005956, 0.000333, -0.001513], std [0.7054, 0.7086, 0.6568, 0.7418, 1.356, 2.424, 0.09317, 0.09378], min [-1, -1, -0.9812, -1, -4.773, -6.141, -0.1991, -0.1995], max [1, 1, 1, 1, 5.034, 6.297, 0.198, 0.1994]
• raw column min [-8.202, -2.947, -4.773, -6.141, -0.1991, -0.1995], max [7.913, 2.943, 5.034, 6.297, 0.198, 0.1994]

Pixels: per channel over all real frames of x / 255: mean [0.2334, 0.3778, 0.5204], std [0.1062, 0.09519, 0.09613].