alignment string | cache_schema_version int64 | cached_episodes int64 | cached_total_frames int64 | cached_valid_transitions int64 | checkpoint_format string | checkpoint_sha256 string | checkpoint_step int64 | dataset_episodes_sha256 string | dataset_fingerprint string | dataset_info_sha256 string | dataset_root string | declared_total_frames int64 | expected_episodes int64 | global_frame_index string | global_frame_index_sha256 string | global_latents string | global_latents_sha256 string | image_key string | image_keys list | latent_dim int64 | latent_dtype string | latent_layout string | latent_normalization dict | missing_episodes list | preprocess_version string | schema_version int64 | source_caches list | source_manifest_sha256 list | status string | transition_frame_stride int64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
row_0_invalid_zero; row_i=z_mu(frame_i-1,frame_i), i>=1 | 1 | 16,000 | 7,636,924 | 7,620,924 | cdlam.stage1.inference | 12ce318e48d5790fb3773f1027edd5a3babe8965027b4957e3ed62d5d19fc386 | -1 | 6f3fa8621c2af67cb732c2101e1e17ba47510560a89b129e3bc1033755eae058 | da232ddfdd1d2ea4434ce8f2d4f4c9f2d742a11ee770b8c12b0306172866f965 | 9b8c5d1a0730a96d81d9c28346626c0669460879aa0e9152a177b2287d5e76a2 | /share/project/sqx/robomme/robomme_16x1000_lerobot_v1 | 7,636,924 | 16,000 | frame_index.npz | 3faaa296e2d1a28617c5b187f40f7a3ef9bdd25c0aa5ea70173bc1a2aba944fa | latents.npy | fb5bb94a6598631a73cafc6c12ae4e6fb200f9c1a6a82a244e40af9eb7bca4df | image+wrist_image | [
"image",
"wrist_image"
] | 64 | float32 | concat(image_z_mu[32],wrist_image_z_mu[32]) | {
"count": 7620924,
"ddof": 0,
"definition": "per_dimension_valid_transition_population",
"epsilon": 0.000001,
"mean": [
-0.025524643000689278,
0.1211475792447839,
-0.22541722494732125,
-0.004860747989105718,
-0.5641124187222499,
-0.34626896221252584,
-0.49602429230981526,
-0.0... | [] | dreamdojo.lam400k.rgb_uint8.crop_4x3.resize_480x640.resize_240x320.v1 | 2 | [
"/share/project/sqx/robomme/standalone/ltm_pi_robomme-resources/artifacts/memory.head32",
"/share/project/sqx/robomme/standalone/ltm_pi_robomme-resources/artifacts/memory.wrist32"
] | [
"1cdd676780f260a5b789895497b6cf9c3bf121e3332ddc8321d09802481cfdc5",
"07fe0cc241b5e01411ea6524a405ff7b32b03ca78757e55c032e0bd6e711fcb4"
] | complete | 1 |
RoboMME-10x DreamDojo Transition Cache
Precomputed DreamDojo LAM transition features for the RoboMME 16-task × 1,000-episode training corpus.
Contents
latents.npy:(7,636,924, 64)float32 latent array. Each row concatenates 32 dimensions from the head camera and 32 dimensions from the wrist camera.frame_index.npz: row-alignedglobal_index,episode_index,frame_index,cache_row, andvalidarrays.manifest.json: cache provenance, alignment contract, normalization statistics, and SHA-256 checksums.
Alignment
Within every episode:
row 0 = zeros and invalid
row i = z_mu(frame_(i-1), frame_i), for i >= 1
The cache contains 16,000 episodes, 7,636,924 frame-aligned rows, and 7,620,924 valid adjacent-frame transitions. The transition stride is 1.
Load the arrays without copying the full cache into RAM:
import numpy as np
latents = np.load("latents.npy", mmap_mode="r")
with np.load("frame_index.npz", allow_pickle=False) as index:
global_index = index["global_index"]
episode_index = index["episode_index"]
frame_index = index["frame_index"]
valid = index["valid"]
Use the hashes and normalization values in manifest.json when validating a download.
- Downloads last month
- 51