Add dataset card / usage README

README.md

@@ -35,28 +35,53 @@ actions recorded over 250 episodes.
 ## Composition
 
 One directory per episode, named `<language_prompt>_<YYYYMMDD_HHMMSS>/`.
-Each directory holds, per autoregressive chunk index `i` (`i = 0, 2, 4, …`):
+Each directory holds three tensor types per autoregressive chunk index `i`:
 
 ```
 <episode>/
-  obs_data_<i>.pt   # raw observation fed to the WAM at chunk i (multi-camera RGB + proprio), pre-VAE
-  latents_<i>.pt    # VAE video latents at chunk i
-  actions_<i>.pt    # predicted action chunk at chunk i
+  obs_data_<i>.pt   # raw observation window fed to the model at chunk i (pre-VAE)
+  latents_<i>.pt    # VAE-encoded video latents at chunk i
+  actions_<i>.pt    # model-predicted action chunk at chunk i
 ```
 
-- `obs_data_<i>.pt` — the raw observation recorded at chunk `i`, before VAE
-  encoding (multi-camera RGB frames + robot proprioceptive state).
-- `latents_<i>.pt` — the VAE-encoded video latents at chunk `i`.
-- `actions_<i>.pt` — the action chunk the model produced at chunk `i`.
-
 Counts: 2079 files of each type (obs_data / latents / actions), across 250
-episodes. The number of chunks per episode varies with task horizon.
+episodes. All tensors are saved with `torch.save`; load with
+`torch.load(path, weights_only=False, map_location="cpu")`. NumPy arrays load
+without torch. Image arrays are `uint8`; latents / actions are `bfloat16`.
+
+### `obs_data_<i>.pt` — `list` of per-frame observation records
+
+Each list element is one frame, a dict in the RoboTwin / aloha-agilex layout:
+
+| key | type / shape | meaning |
+|---|---|---|
+| `observation.images.cam_high` | `uint8 [240, 320, 3]` | head camera RGB (HWC) |
+| `observation.images.cam_left_wrist` | `uint8 [240, 320, 3]` | left-wrist camera RGB |
+| `observation.images.cam_right_wrist` | `uint8 [240, 320, 3]` | right-wrist camera RGB |
+| `observation.state` | `float64 [14]` | proprioceptive state (dual 7-DoF arms) |
+| `task` | `str` | language instruction |
+
+The list is **cumulative** — it grows as the rollout proceeds (4 records at
+chunk 0, 8 at chunk 2, … i.e. the full observation history up to chunk `i`).
+
+### `latents_<i>.pt` — `bfloat16 [1, 48, 2, 24, 20]`
+
+`[batch, latent_channels=48, temporal_frames=2, H_lat=24, W_lat=20]` — the WAN
+VAE-encoded video latent for the 2 frames generated at chunk `i`.
+
+### `actions_<i>.pt` — `bfloat16 [1, 30, 2, 16, 1]`
+
+`[batch, action_dim=30, temporal_frames=2, steps=16, 1]` — the model-predicted
+action chunk for chunk `i` (fixed shape per chunk, not cumulative).
 
-All tensors are saved with `torch.save`; load with
-`torch.load(path, weights_only=False, map_location="cpu")`.
+### Chunk indexing
 
-Episode directories are flattened at the repo root (the upload tool did not
-preserve a parent prefix); each prompt+timestamp directory name is unique.
+Chunk indices are **even and contiguous** (`0, 2, 4, 6, …`): the WAN VAE
+temporally downsamples 2:1, so each autoregressive step emits 2 latent frames,
+and `<i>` is the frame-start id of that step (hence the step of 2). The set of
+indices present per episode runs `0 … 2·(num_chunks − 1)`; episode length varies
+with task horizon. Episode directories are flattened at the repo root (the
+upload tool did not preserve a parent prefix); each name is unique.
 
 ## Download