value-model-benchmark-training-data / data /simulation /one-shot /annotations /subtasks /load_subtasks.py
| """Read supplemental simulation One-Shot labels beside a LeRobot v3.0 dataset. | |
| The sidecar labels the active semantic subtask, not reward or success. Semantic | |
| boundary adjustment bands are independent of LeRobot video-loading tolerance. | |
| Requires only the Python standard library and PyArrow. | |
| """ | |
| from pathlib import Path, PurePosixPath | |
| import hashlib | |
| import json | |
| import pyarrow.parquet as pq | |
| class SubtaskAnnotations: | |
| def __init__(self, dataset_root, verify_manifest=True): | |
| self.dataset_root = Path(dataset_root) | |
| self.annotation_root = self.dataset_root / "annotations" / "subtasks" | |
| self.metadata = json.loads((self.annotation_root / "metadata.json").read_text()) | |
| if verify_manifest: | |
| manifest = json.loads((self.annotation_root / "manifest.json").read_text()) | |
| for entry in manifest["files"]: | |
| name = PurePosixPath(entry["path"]) | |
| if name.is_absolute() or ".." in name.parts: | |
| raise ValueError("Invalid manifest-relative file path") | |
| payload = (self.annotation_root / name).read_bytes() | |
| if len(payload) != entry["bytes"] or hashlib.sha256(payload).hexdigest() != entry["sha256"]: | |
| raise ValueError("Annotation manifest checksum mismatch") | |
| info_path = self.dataset_root / "meta" / "info.json" | |
| if info_path.exists(): | |
| info = json.loads(info_path.read_text()) | |
| for key in ("fps", "total_episodes", "total_frames"): | |
| if info.get(key) != self.metadata["dataset"][key]: | |
| raise ValueError("Dataset/annotation metadata mismatch: " + key) | |
| if info.get("codebase_version") != "v3.0": | |
| raise ValueError("Expected a LeRobot v3.0 dataset") | |
| self.table = pq.read_table(self.annotation_root / "frame_labels.parquet") | |
| self.by_frame = {} | |
| global_indices = set() | |
| for row in self.table.to_pylist(): | |
| key = (row["episode_index"], row["frame_index"]) | |
| if key in self.by_frame or row["global_index"] in global_indices: | |
| raise ValueError("Duplicate annotation frame/global index") | |
| self.by_frame[key] = row | |
| global_indices.add(row["global_index"]) | |
| if len(self.by_frame) != self.metadata["dataset"]["total_frames"]: | |
| raise ValueError("Annotation frame count mismatch") | |
| if global_indices != set(range(self.metadata["dataset"]["total_frames"])): | |
| raise ValueError("Annotation global index grid mismatch") | |
| for episode in self.metadata["episodes"]: | |
| ep = episode["episode_index"] | |
| for frame in range(episode["length"]): | |
| row = self.by_frame.get((ep, frame)) | |
| if row is None or row["global_index"] != episode["dataset_from_index"] + frame: | |
| raise ValueError("Annotation episode/frame/global index mismatch") | |
| def label(self, episode_index, frame_index, global_index=None): | |
| row = self.by_frame[(int(episode_index), int(frame_index))] | |
| if global_index is not None and row["global_index"] != int(global_index): | |
| raise ValueError("LeRobot sample index does not match annotation") | |
| return dict(row) | |
| def for_sample(self, sample): | |
| """Attach explicitly to a scalar LeRobotDataset sample, before collation.""" | |
| return self.label(sample["episode_index"], sample["frame_index"], sample["index"]) | |
| def load_subtasks(dataset_root, verify_manifest=True): | |
| return SubtaskAnnotations(dataset_root, verify_manifest=verify_manifest) | |