---

name: latency-submissions description: "Guides real-time Waymo Open Dataset latency challenge submissions, wod_latency_submission modules, numpy detection inputs and outputs, Docker image sources, and latency result conversion." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Latency Submissions

Use this sub-skill when the task is about the Waymo real-time 2D or 3D detection latency challenge: validating a wod_latency_submission module, choosing DATA_FIELDS, preparing pre-extracted numpy inputs, interpreting boxes/scores/classes outputs, packaging a challenge Docker image source, or converting latency evaluator outputs toward an Objects proto workflow.

Read:

• references/workflows.md for the module contract, evaluator flow, Docker image-source requirements, and result-conversion flow.
• references/data-formats.md for latency numpy field names, previous-frame suffixes, result-directory layout, and 2D versus 3D output shapes.
• references/troubleshooting.md for import errors, missing .npy inputs, output length mismatches, 2D camera restrictions, and objects skipped during conversion.

Useful bundled checks:

• scripts/validate_latency_submission.py imports a user-provided module or file, creates tiny fake numpy arrays for its DATA_FIELDS, calls initialize_model() and run_model(...), and validates the returned arrays without importing the Waymo package.
• scripts/make_latency_fixture.py creates or checks a tiny latency result directory and writes a JSON summary without requiring the Waymo package.

Route ordinary metric accuracy to metrics-evaluation, raw Frame conversion to dataset-utils, and generic Docker/PyPI package build work to repo-build-test.