physicalai-train-exporting-and-validating

Exporting and Validating Studio Policies

Export lives in library/src/physicalai/export/: backends.py (the ExportBackend enum — onnx, openvino, torch, executorch — plus per-backend parameter classes) and mixin_policy.py (ExportablePolicyMixin, which gives policies export(output_dir, backend=...)). The Python API is primary library behavior; the CLI entry library/src/physicalai/cli/export.py must preserve the same artifact contract. Studio owns export; Runtime owns loading.

Workflow

1. Identify the inputs: source policy class (e.g. physicalai.policies.ACT), .ckpt path, target backend, and the Runtime loader behavior expected for that backend.
   • Done when: all four are pinned before touching code.
2. Pick the route and keep both consistent — they must produce the same artifact:
   • Python: policy.export(output_dir, backend=ExportBackend.ONNX).
   • CLI: physicalai export --policy physicalai.policies.ACT --ckpt_path model.ckpt --backend onnx --output_dir ./export.
3. Read backend constraints before editing generic code. See the backend reference for the target (references/<backend>.md). Do not generalize a fix across backends without checking each.
4. Export, then validate numerical parity against the Torch policy path on representative inputs. Parity proves correctness.
   • Done when: max abs/rel diff on sample inputs is within the family's tolerance, or the divergence is understood and documented.
5. Validate artifact structure and metadata against references/export-contract.md.
   • Done when: the expected model file and metadata files exist, and input/output/feature names match Runtime preprocessing.
6. Confirm the Runtime path. For deployment-bound artifacts, verify Runtime can auto-detect (by extension) or explicitly load the backend via InferenceModel(...).

Validation loop

Run export → validate → fix → repeat until both parity and structure pass:

# from library/
physicalai export --policy <ClassPath> --ckpt_path <model.ckpt> --backend <backend> --output_dir ./export
uv run pytest tests/unit/export -k <backend>

For API-facing changes, add or run an equivalent Python script/test that loads the checkpoint, calls policy.export("./export-api", backend=ExportBackend.<BACKEND>), and compares artifact metadata with the CLI output.

Treat parity (correctness) and latency/warmup (deployment viability) as separate checks; passing one does not imply the other.

Backend notes

• onnx / openvino — deployment-oriented; Runtime core ships adapters, so artifacts load when deps are installed.
• torch — development/debugging; only claim deployment support when a matching Runtime adapter is installed and documented.
• executorch — optional, dependency-sensitive, edge/mobile; Runtime core ships no adapter in this package. Treat as available only with a documented companion distribution.

Required checks

• Export directory contains the expected backend model file and metadata files.
• Metadata names inputs/outputs/features consistently with Runtime preprocessing and action-chunk semantics.
• Python API export and CLI export produce equivalent artifact structure and metadata.
• Backend-specific dependencies are imported lazily or guarded with clear install guidance.
• Do not add a backend to user-facing docs unless Runtime can load it in-package or via a documented companion.
• CLI docs (library/docs/how-to/export/) and Python API examples stay consistent.

References

• references/export-contract.md — artifact requirements shared with Runtime (keep synchronized; CI should fail on divergence).
• references/onnx.md, references/openvino.md, references/torch.md, references/executorch.md — per-backend constraints.