1.9 KiB
1.9 KiB
name, description
| name | description |
|---|---|
| deploy | Export and validate research-only ONNX inference bundles with matching evaluation provenance, numerical parity and explicit hardware exclusions. |
Package a policy for research inference
Read Deployment contract. Complete train/controls evidence first. This is not a hardware deployment skill: no motor enable, robot transport or firmware operation.
Procedure
- Select a committed checkpoint from a completed run. Inspect its training curves, conformance report, fixed-protocol evaluation, censoring, CSV and video. Export can package a poor policy; it does not certify quality.
- Finish baseline evaluation/export before editing source. Restore historical sources/dependencies in a separate trusted environment when needed; do not bypass provenance checks.
- Run
./scripts/uh export --checkpoint CHECKPOINT --evaluation REPORT --output NEW_BUNDLE. The evaluator must match checkpoint, experiment and source identity. ONNX parity uses actual rollout observations, not just random input vectors. - Run
./scripts/uh verify-bundle BUNDLE --inference. Inspect observation fields, normalizer, joint order, action scale/offset/clip, PD metadata, units and control timing. - Review asset/code licenses and manifests for private paths before publishing artifacts. Source archives are trusted restoration material, not automatically executable downloads.
Failure handling and handoff
Hash mismatch, source drift, unsupported action schema or failed parity is a blocker. Do not
edit hashes or tolerance to turn an unknown failure into success. Hand over bundle/report paths,
measured parity, training/evaluation limitations and the fact that hardware_approved remains
false. Physical use additionally needs validated estimation, transport, joint mapping, watchdog,
limits, emergency stop and operator-approved testing; those are absent here.