# Robotics dataset QA report template

Version: 1.0.0. Updated: 2026-11-10. Maintainer: Avala Solutions Engineering. Language: English.

Blank report template, not measured results, a customer case study or an Avala quality guarantee. Choose applicable checks and agree on definitions before evaluation. All result and acceptance fields are intentionally empty. Keep confidential recordings, identifiers and credentials out of public copies.

## Release and review scope

| Required record | Value / evidence |
| --- | --- |
| Report owner, reviewer and date | |
| Release ID, version and manifest reference | |
| Data product, embodiment and intended learning task | |
| Output contract and taxonomy versions | |
| Population submitted and units | |
| Review sample and selection method | |
| Excluded, inaccessible and unreviewed data | |
| Reference sample and its provenance | |
| Reviewer qualification and independence | |
| Agreed acceptance criteria and decision owner | |

Do not treat a target as an observed result. Keep full-population checks separate from sampled review. A result from one sensor configuration, task family or environment does not establish quality in another.

## Task-specific checks

Select checks appropriate to the program. These rows are candidate review dimensions, not a claim that every modality or metric is supported by Avala.

| Dimension | Definition and tolerance | Reference / method | Numerator | Denominator | Result / uncertainty | Agreed threshold | Decision |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Geometry / class / attribute correctness | | | | | | | |
| Missed objects and critical errors | | | | | | | |
| Track continuity / identity switches | | | | | | | |
| Temporal action boundaries | | | | | | | |
| Episode technical validity | | | | | | | |
| Required observations/actions/state completeness | | | | | | | |
| Task success/failure/recovery review | | | | | | | |
| Calibration residuals / validation conditions | | | | | | | |
| Time alignment / coordinate frame consistency | | | | | | | |
| Export fidelity / downstream loader result | | | | | | | |

For a rate, state the counted event, eligible population, denominator and exclusions. A zero or unavailable denominator is **not assessed**, not zero error. For timing or geometry tolerances, state units and the reference setup. Record confidence intervals or other uncertainty when justified by the sampling method; do not invent statistical precision.

## Disagreement and adjudication

| Item | Evidence / decision |
| --- | --- |
| Reviewers and number of independent reviews | |
| Disagreement definition and measurement | |
| Adjudication owner and resolution method | |
| Taxonomy ambiguities and versioned corrections | |
| Unresolved cases and disposition | |
| Re-review scope after correction | |

Keep original review outcomes and adjudicated decisions distinguishable. If the taxonomy changes during review, identify affected records and whether earlier decisions were reassessed.

## Yield, coverage and rework

| Measure | Definition / denominator | Observed value | Evidence and limits |
| --- | --- | --- | --- |
| First-pass acceptance | | | |
| Final accepted quantity | | | |
| Rejected quantity by reason | | | |
| Rework attempts and review effort | | | |
| Environment / task / condition coverage | | | |
| Duplicate or excluded data | | | |
| Turnaround | | | |

Report both first-pass and final acceptance; do not erase rejected attempts from the history. State whether quantities are episodes, sequences, labels or hours. Use the coverage brief to report missing cells. Commercial treatment of rework must be agreed separately; this report does not set who pays.

## Model evaluation, if performed

Keep annotation QA separate from downstream model performance. Record training/evaluation versions, fixed controls, held-out split construction, leakage checks, metric definitions, baseline and uncertainty. If no controlled model evaluation occurred, state **not evaluated**. Loader compatibility alone does not demonstrate policy improvement.

## Release decision and limits

| Decision | Owner / evidence |
| --- | --- |
| Accept, reject or hold for further review | |
| Unresolved critical issues and containment | |
| Approved uses and excluded uses | |
| Required corrections or recollection | |
| Next review / expiry of this assessment | |

An empty report is not acceptance. Retain evidence sufficient to reproduce the result under authorized access. Avoid exposing customer recordings or private identifiers in a public version.

## Publication of completed results

Public use of completed results approved: **No** until explicitly recorded.

Approver, approved claims/artifacts, customer permission, measurement scope and review date:


## Related resources

- [Robotics evaluation checklist](https://avala.ai/usecases/robotics)
- [Labeling specifications](https://avala.ai/platform/labeling-capabilities)
- [Benchmark worksheet](https://avala.ai/media/templates/robotics-benchmark-worksheet-v1.md)
- [Discuss labeling](https://avala.ai/demo?intent=labeling)