Apairo

One numpy-in / numpy-out API for robotics sensor datasets.

What is apairo?

apairo is a unified loader for robotics sensor datasets — lidar, cameras, poses, IMU, labels — with one chainable API across synchronous (KITTI-style) and asynchronous (multi-rate) layouts. No bespoke glob + np.load loaders, no hand-written timestamp matching: everything is a dataset you filter, select, cache, synchronize, concat and feed straight to a PyTorch DataLoader.

• numpy in, numpy out — ds[i].data["lidar"] -> np.ndarray; convert to torch/tf at the edge, never inside the dataset.
• everything is a dataset — filter / select / cache / join / concat / synchronize / split / transform all return chainable, lazy views.
• .apairo is the source of truth on disk — channels (raw + derived) are declared in a sidecar; expensive preprocess output is persisted and reloaded transparently.

import apairo

ds = (apairo.RawDataset("/data/mission", keys=["lidar", "image", "gps"])
        .synchronize(reference="lidar", method="nearest", tolerance=0.05)
        .filter(lambda s: s.data["lidar"].shape[0] > 1000))

ds[0].data          # {"lidar": (N, 4), "image": (H, W, 3), "gps": (3,)}
# an apairo dataset *is* a torch Dataset -> DataLoader(ds, collate_fn=...), no adapter

Point RawDataset at a directory and it loads — no code. When a channel's clock lives in its filenames, declare it right there and apairo reads it in memory, never writing into your data. See Bring your own dataset.

The ecosystem

Mechanisms live in the core; collections live in satellites. The core never gains a dependency beyond numpy + PyYAML.

Getting started

pip install apairo          # Python >= 3.11
pip install apairo[vision]  # optional: image loading (Pillow)

• Documentation — quickstart, the .apairo schema, synchronizing async sensors, preprocessing.
• Bring your own dataset — YAML profiles, thin subclasses, filename-encoded keys.
• apairo-robotics/apairo — source, issues, contributing.