Loader Examples¶

RoboMetrics loaders are small helpers for turning local trajectory files into arrays that can be passed to metrics or Evaluator.evaluate_dataset().

Directory-Level Loading¶

Use load_trajectory_dir() when you want every supported trajectory file in a directory as a filename-keyed dictionary:

from robometrics import load_trajectory_dir

trajectories = load_trajectory_dir("rollouts/predictions")
for name, trajectory in trajectories.items():
    print(name, trajectory.shape)

The helper loads .npy, .npz, .csv, and .json files and skips unsupported extensions. CSV loading requires robometrics[io] because it uses pandas.

Dataset-Level Evaluation¶

Use trajectories_to_dataset() when you have matched prediction and reference files. It preserves file order, so sort the paths or use your own manifest when the pairing matters.

from pathlib import Path

from robometrics import Evaluator
from robometrics.loaders import trajectories_to_dataset

pred_dir = Path("rollouts/predictions")
gt_dir = Path("rollouts/references")

predictions, ground_truths = trajectories_to_dataset(
    sorted(pred_dir.glob("*.csv")),
    sorted(gt_dir.glob("*.csv")),
)

result = Evaluator().evaluate_dataset(
    predictions=predictions,
    ground_truths=ground_truths,
    metrics=["ade", "fde"],
    thresholds={"ade": 0.25, "fde": 0.5},
)

print(result.to_markdown())

For a complete runnable example that creates temporary CSV/JSON files, loads a directory, evaluates a matched dataset, and prints the schema version, see examples/dataset_loader_evaluation.py.