# Robotics data workflow benchmark worksheet

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

Blank planning template, not a customer result or an Avala commercial commitment. Keep confidential recordings, identifiers and credentials out of public copies.

Blank evaluation worksheet. Complete the output contract first. No default quality target, price, throughput or improvement is implied. Use the same reference data and conditions for each workflow.

| Required record | Agreed value or receipt |
| --- | --- |
| Evaluation owner and date | |
| Intended training task and known failure modes | |
| Input slice IDs, selection method and exclusions | |
| Output contract/version and schema | |
| Reference annotations, provenance and independent reviewer | |
| Baseline workflow and candidate workflow versions | |
| Blinding/randomization method, where applicable | |
| Held-out evaluation population and leakage controls | |

## Integration feasibility

| Check | Baseline | Candidate | Evidence |
| --- | --- | --- | --- |
| Recording ingest, supported channels and rejected inputs | | | |
| Coordinate frames, calibration and temporal alignment | | | |
| Commanded action vs measured state vs estimated pose | | | |
| Annotation/review to export without undeclared loss | | | |
| Actual training loader accepts the delivered format/version | | | |
| Ownership, portability, access and retention requirements | | | |

## Accepted output and economics

| Measure | Definition / denominator | Baseline | Candidate |
| --- | --- | --- | --- |
| Submitted quantity | Episodes, sequences or hours; choose one | | |
| Accepted quantity | After technical and task/outcome QA | | |
| Coverage | Environment/condition/task/embodiment cells | | |
| Critical errors | Predefined categories and reference sample | | |
| Temporal identity | Track switches, interruptions and assessment window | | |
| Review disagreement | Reviewer population and adjudication method | | |
| Rework | Attempts, time and rejected-output disposition | | |
| All-in cost | Integration, calibration, annotation, review and rework | | |
| Cost per accepted unit | Total included cost / accepted quantity | | |
| Turnaround | Same start/end event definitions | | |

## Model effect and next iteration

Record training code/version, compute/data controls, held-out conditions, metrics, uncertainty and limitations. Keep integration success separate from demonstrated model improvement. A small pilot is not fleet-scale evidence.

Trace one observed failure to a reviewed data correction, a versioned release, a loader receipt and the next evaluation. Record what changed and what remained unchanged. Decide the next scope using the agreed evidence, not a vendor-specific winning threshold.

## Publication of completed results

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

Approver, permitted claims/artifacts, customer consent, measurement scope and expiry/review date:


## Related resources

- [Robotics workflow and evaluation checklist](https://avala.ai/usecases/robotics)
- [Labeling specifications](https://avala.ai/platform/labeling-capabilities)
- [Discuss a program](https://avala.ai/demo)