What is Avala?

Here is a familiar robotics problem: a robot did something unexpected, and the evidence is spread across camera recordings, LiDAR, joint states, and calibration files. Before the next training run, the team needs to understand the event, decide what to label, and review the resulting data.

Avala is Fleet Learning Infrastructure for Physical AGI. It connects sensor data, visualization, human and AI labeling, quality review, and training-data delivery. The aim is a traceable path from a recording to the data a model learns from, with field failures informing the next round of work.

Why the labels matter

A recording can show what happened without explaining what a model should learn. A perception task may need object identities across frames; a manipulation task may need contact phases or keypoints. The right labels depend on the model, the sensors, and the evaluation you are trying to improve.

Quality requires an explicit specification and review. A tracking identifier, a projected cuboid, or an interpolated label still needs to be checked against the source data. Calibration errors and ambiguous scenes do not disappear because the annotation tool shares a timeline.

What this course covers

This course explains the workflow and the decisions at each stage. It is a conceptual guide, with selected SDK and CLI examples for your own workspace; it does not supply a sample recording or run a training experiment.

1. Ingest and visualize — prepare sensor data and inspect its timing and geometry.
2. Annotate — create model-assisted labels and review them against your task specification.
3. Curate — choose the examples, quality states, and metadata your training set needs.
4. Train — export reviewed data and connect it to your existing training pipeline.
5. Close the loop — use evaluation and field failures to decide what data to collect or label next.

Datasets, projects, tasks, and exports are the working pieces. Each lesson introduces the relevant concepts and points you to the developer documentation for implementation details.

Connect your workspace

For the examples, install the Python SDK with its CLI dependencies:

pip install "avala[cli]"

Create an API key using the quickstart, and provide it through the AVALA_API_KEY environment variable. Run examples in a workspace you are authorized to access. Listing datasets is a read-only first check:

from avala import Client

with Client() as client:
    for dataset in client.datasets.list(limit=10):
        print(dataset.uid, dataset.name)

An empty result means there are no visible datasets on this page of results; it does not create a sample dataset. The remaining lessons explain how to work with your own data.

Next: The Ground-Truth Loop →