Add Wall-X serving and Turtle2 TCP WebSocket bridge

scripts/README.md

@@ -22,6 +22,47 @@ python scripts/fake_inference.py \
   --train-config-path /path/to/config.yml
 ```
 
+## Converting LeRobot v3 data for the pinned v2.1 loader
+
+`convert_lerobot_v3_to_v21.py` creates a separate, episode-sharded v2.0
+dataset compatible with Wall-X's pinned LeRobot v2.1 loader. It converts both
+the Parquet data and video streams; the source directory is never changed.
+
+```bash
+python scripts/convert_lerobot_v3_to_v21.py \
+  --source /path/to/libero_v3 \
+  --output /path/to/libero_v21 \
+  --dry-run
+
+python scripts/convert_lerobot_v3_to_v21.py \
+  --source /path/to/libero_v3 \
+  --output /path/to/libero_v21
+```
+
+Use `--max-episodes 1` only to test the conversion pipeline. It creates a
+subset and must not be used for the full training run.
+
+## Merging compatible LeRobot v2.1 datasets
+
+`merge_lerobot_v21_datasets.py` combines compatible episode-sharded v2.1
+datasets into one root for a single training run. It renumbers global task,
+episode, and frame indices. On one filesystem videos are hard-linked, so the
+merge does not duplicate MP4 contents. The sources are never modified.
+
+```bash
+python scripts/merge_lerobot_v21_datasets.py \
+  --sources /path/to/libero_spatial /path/to/libero_object /path/to/libero_goal /path/to/libero_10 \
+  --output /path/to/libero_all_v21 \
+  --dry-run
+
+python scripts/merge_lerobot_v21_datasets.py \
+  --sources /path/to/libero_spatial /path/to/libero_object /path/to/libero_goal /path/to/libero_10 \
+  --output /path/to/libero_all_v21
+```
+
+Pass `--copy-videos` only when the output is on another filesystem and cannot
+use hard links; it duplicates the video storage.
+
 ## LIBERO evaluation
 
 `run_libero.sh` is a small shell wrapper around `infer_libero.py`. It requires