[Fix] Update dataset conversion for InternData-N1 VLN-PE v0.5 dataset format (#288)

* update lerobot dataset

* add compatibility doc

* remove chinese comment

---------

Co-authored-by: wangyukai <wangyukai@pjlab.org.cn>

Author: kew6688

docs/compatibility.md

@@ -0,0 +1,13 @@
+# Compatibility
+
+## v0.3.1
+
+### InternData-N1 update to v0.5
+
+The InternData-N1 VLN-PE trajectory training dataset has been upgraded from `v0.1` to `v0.5`. This update introduces minor structural changes in the dataset layout and updates the LeRobot-to-LMDB conversion logic to match the new `v0.5` data structure.
+
+The training pipeline now uses the new key name:
+- `instruction_text` → `task`
+
+The updated conversion logic is **not compatible** with InternData-N1 `v0.1`.
+

internnav/dataset/cma_lerobot_dataset.py

@@ -6,7 +6,7 @@
 
 from internnav.dataset.base import BaseDataset, ObservationsDict, _block_shuffle
 from internnav.model.utils.feature_extract import extract_instruction_tokens
-from internnav.utils.lerobot_as_lmdb import LerobotAsLmdb
+from internnav.utils.loader import LerobotAsLmdb
 
 
 class CMALerobotDataset(BaseDataset):
@@ -38,8 +38,9 @@ def __init__(
         self.camera_name = self.config.il.camera_name
 
         self.lerobot_as_lmdb = LerobotAsLmdb(self.lerobot_features_dir)
-        self.lmdb_keys = self.lerobot_as_lmdb.get_all_keys()
+        self.lmdb_keys = self.lerobot_as_lmdb.get_all_keys(allow_scan_list=['r2r']) # r2r / r2r_aliengo / r2r_flash
         self.length = len(self.lmdb_keys)
+        print(f"total keys in traj_data: {len(self.lmdb_keys)}")
 
         # For CMA-CLIP
         self.use_clip_encoders = False

internnav/dataset/rdp_lerobot_dataset.py

@@ -26,7 +26,7 @@
 from internnav.model.basemodel.LongCLIP.model import longclip
 from internnav.model.utils.feature_extract import extract_instruction_tokens
 from internnav.utils.geometry_utils import get_delta, normalize_data, to_local_coords
-from internnav.utils.lerobot_as_lmdb import LerobotAsLmdb
+from internnav.utils.loader import LerobotAsLmdb
 
 
 def _convert_image_to_rgb(image):
@@ -103,8 +103,9 @@ def __init__(
         self.to_pil = ToPILImage()
         self.image_processor = _transform(n_px=224)  # copy from clip-long
         self.lerobot_as_lmdb = LerobotAsLmdb(self.lerobot_features_dir)
-        self.lmdb_keys = self.lerobot_as_lmdb.get_all_keys()
+        self.lmdb_keys = self.lerobot_as_lmdb.get_all_keys(allow_scan_list=['r2r']) # r2r / r2r_aliengo / r2r_flash
         self.length = len(self.lmdb_keys)
+        print(f"total keys in traj_data: {len(self.lmdb_keys)}")
 
         self.start = 0
         self.end = self.length
@@ -192,7 +193,7 @@ def _load_next(self):  # noqa: C901
                 episodes_in_json = data_to_load['episodes_in_json']
 
                 instructions = [
-                    episodes_in_json[ep_idx]['instruction_text'][: self.config.model.text_encoder.max_length]
+                    episodes_in_json[ep_idx]['task'][: self.config.model.text_encoder.max_length]
                     for ep_idx in range(len(episodes_in_json))
                 ]