> ## 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.

# Coverage

> Coverage cells, concentration limits, overlap checks, the scene-realism standard and the coverage report in every delivery

More hours of the same room and the same motion add little. Coverage is how Avala plans and proves variety across the dimensions that change what a model learns.

## Coverage cells

A coverage cell is one combination of six dimensions.

| Dimension | Examples |
| - | - |
| Environment | Galley kitchen, laundry room, bedroom, garage, hotel room |
| Mechanism | Pull-out dishwasher rack, French-door fridge, top-load washer, cordless kettle base |
| Material | Wet ceramic, flexible bag, folded cotton, rigid plastic tub |
| Interaction primitive | Open, grasp, pour, place, fold, insert, latch |
| Failure mode | Slip and regrasp, missed placement, collision with an obstacle, none |
| Embodiment | Human egocentric, single arm, bimanual arm, mobile manipulator |

Acceptance records the cell values of each capture. The MCP tool `get_campaign_acceptance_coverage` reports which values are under-covered across accepted captures. A brief sets target accepted hours per cell. A [capability experiment](/docs/fleet-learning/capability-experiments.md) lists the cells it targets.

Avala plans collection from a public [catalog of 175 workflows](https://avala.ai/coverage/catalog) in 26 task families.

## Concentration limits

A brief sets the maximum share of accepted hours that any one value can take. Typical limits apply to:

* one location
* one operator
* one appliance or mechanism unit
* one coverage cell
* low-variety actions, such as repeated wiping

Avala measures concentration on every delivery. A sequence that would push a value past its limit is held back and reported.

## Overlap checking

Avala fingerprints the video content of every sequence. Overlap checking runs in three passes:

1. **Exact files.** Identical files are caught by hash.
2. **Content fingerprints.** Re-encoded, trimmed or cropped copies of the same footage are caught by fingerprint match.
3. **Similarity review.** Near-identical scenes from the same place and setup are flagged for a reviewer.

The brief states what each delivery is compared against: the delivery itself, earlier deliveries, or reference footage you provide. Fingerprints are versioned. The coverage report names the version and the scope that were searched.

## Scene-realism standard

Avala collects real household work in real rooms. The standard has four rules.

* **Keep ordinary distractors.** Leave unrelated, safe items in view, such as mail, a phone or adjacent groceries. Nobody clears a whole table to move one bowl.
* **Capture meaningful work.** Use natural pace and a real household goal. Slow motion for its own sake, repeated wiping loops, artificial spills and staged piles fail the standard.
* **Vary the whole setting.** Spread collection across homes, layouts, operators and appliances. Swapping one utensil or relabeling the same action does not make a new environment.
* **Allow purposeful arrangement.** A real request, such as resetting a desk, is valid. Record the starting state, the goal and the result. Do not create disorder to start the task.

Distractors stay in view only when consent, privacy and safety allow it.

## The coverage report

Each delivery ships with a coverage report. The brief fixes its fields, definitions and thresholds before collection starts.

| Section | Contents |
| - | - |
| Environment and people | Pseudonymous location and collector IDs. Hours and sequences per ID. Room types, layouts and concentration against limits. |
| Appliance evidence | Appliance category, format and controls observed on site. Coverage by appliance and setting. |
| Task and action coverage | Workflows. Objects, subtasks and interaction primitives. Starting and end states. The share of low-variety actions. |
| Coverage cells | Accepted hours per cell against the target, with empty and thin cells listed. |
| Naturalness and motion | Review of pace, purposeful movement, repeated loops and staged conditions. Accepted and rejected examples for calibration. |
| Overlap and similarity | Exact-file, fingerprint and similarity results. The fingerprint version and the comparison scope. |
| Rights and acceptance | Capture provenance and consent records. Accepted hours, rejected hours and [rejection reasons](/docs/fleet-learning/accepted-hour.md#rejection-reasons). Metadata schema version. |

Before delivery, Avala compares the measured mix with the brief. The report lists over-represented locations, operators, appliances and actions next to the missing cells.

## Read coverage with the MCP server

The MCP tool `get_campaign_acceptance_coverage` returns accepted-capture counts per axis value for a capture campaign project. Pass `axes` as a comma-separated list to choose the axes. Values with low counts show where the campaign is thin. See [Use Avala with AI](/docs/getting-started/mcp.md) to connect your assistant.