> ## Documentation Index
> Fetch the complete documentation index at: https://avala.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Fleet learning

> How Avala turns your fleet's failures into the data, evaluation and releases your next model needs

Avala is Fleet Learning Infrastructure for Physical AGI. Every fleet should learn from its own experience. Avala sits in the infrastructure layer under your model. You train the model and you own it. Avala runs the loop around it: evaluation, failure diagnosis, collection, validation and release.

This section documents the fleet learning program. It covers the loop, the records Avala keeps, the schemas Avala delivers and the reports you receive.

## The loop

Each cycle starts from what your fleet does in the real world and ends with a release you can train on.

**1. Observe**

Your robots run. Rollouts, interventions and failures flow into Avala as fleet recordings and events. See [Failure ingestion](/docs/fleet-learning/failure-ingestion.md).

**2. Diagnose**

Avala segments the rollouts, clusters the failures into a failure taxonomy and writes a gap report.

**3. Evaluate**

Avala scores your current model version on held-out real-world evaluation scenes. See [Evaluation](/docs/fleet-learning/evaluation.md).

**4. Acquire**

Avala collects the data the gap report calls for. Each batch targets specific [coverage cells](/docs/fleet-learning/coverage.md) and counts only [accepted hours](/docs/fleet-learning/accepted-hour.md).

**5. Train**

You train your model on the release. Training stays on your side.

**6. Validate**

Avala re-evaluates the new model version on the same held-out scenes and reports the change per failure mode.

**7. Release**

Avala publishes a versioned, rights-cleared dataset release with [sequence metadata](/docs/fleet-learning/sequence-metadata.md) and [training exports](/docs/fleet-learning/exports.md).

**8. Observe**

The new model goes back to your fleet. Its rollouts start the next cycle.

## Who does what

| Step | You | Avala |
| - | - | - |
| Observe | Run the fleet. Send rollouts, logs and interventions. | Ingests recordings and events. Checks sensor integrity. |
| Diagnose | Review the gap report. Set priorities. | Segments rollouts, clusters failures, writes the gap report. |
| Evaluate | Provide model versions or rollouts on the evaluation scenes. | Maintains held-out scenes, scores runs, reports metrics. |
| Acquire | Approve the brief. | Designs the capture campaign, collects, reviews and accepts sequences. |
| Train | Train and own the model. | Delivers the release in the formats your training stack reads. |
| Validate | Run the new model version. | Re-evaluates and reports the capability change. |
| Release | Decide what ships to the fleet. | Versions the dataset, records provenance and consent, publishes the release. |

You decide what the model learns and when it ships. Avala supplies the evidence and the data for each decision.

## How an engagement runs

Every engagement follows the same five steps. Each step produces an artifact you can inspect.

1. **Failure.** You send rollouts where the model failed or needed help. Avala reads them as fleet recordings and events.
2. **Gap report.** Avala groups the failures by failure mode and names the coverage cells that are missing or thin.
3. **Brief.** Avala turns the gap report into a collection brief. The brief lists target coverage cells, concentration limits, acceptance criteria and the evaluation cohort. You approve it.
4. **Batch.** Avala collects and reviews the batch. Each delivery ships with a [coverage report](/docs/fleet-learning/coverage.md#the-coverage-report) and per-sequence metadata.
5. **Re-evaluation.** Avala scores your next model version on the held-out scenes. The [capability report](/docs/fleet-learning/evaluation.md#the-capability-report) shows what changed and sets the next allocation.

Gap reports, briefs and capability reports are delivered as part of the program. You can read the objects behind them, such as fleet recordings, capture campaigns, datasets and slices, with the SDKs and the MCP server.

Avala tracks each cycle as a [capability experiment](/docs/fleet-learning/capability-experiments.md). The record links the model version, the failure, the data and the outcome.

## In this section

> **[Capability experiments](/docs/fleet-learning/capability-experiments.md)**
>
> The record that ties a failure to the data collected for it and the measured outcome.

> **[Sequence metadata schema](/docs/fleet-learning/sequence-metadata.md)**
>
> The outcome label on every sequence, how technical validity is reported, and how to read and export both.

> **[Accepted hour](/docs/fleet-learning/accepted-hour.md)**
>
> The unit Avala delivers, and the rejection reasons Avala reports.

> **[Coverage](/docs/fleet-learning/coverage.md)**
>
> Coverage cells, concentration limits, overlap checks and the coverage report.

> **[Evaluation](/docs/fleet-learning/evaluation.md)**
>
> Held-out scenes, scoring, metrics and the capability report.

> **[Failure ingestion](/docs/fleet-learning/failure-ingestion.md)**
>
> Send rollouts and failures with the SDK, CLI and MCP server.

> **[Exports](/docs/fleet-learning/exports.md)**
>
> LeRobot import and export, and what ships with each release.