physicalai-runtime-configuring-inference-pipeline

Configuring the Inference Pipeline

Pipeline order: observation → preprocessors → runner → postprocessors → action output. See docs/how-to/inference/configure-pre-post-processing.md.

Core code:

• src/physicalai/inference/component_factory.py — ComponentRegistry, instantiate_component, _MAX_COMPONENT_DEPTH.
• src/physicalai/inference/model.py — builds processor chains from manifest specs.
• Built-ins under preprocessors/ and postprocessors/; runners under runners/.

Workflow

1. Read the manifest slice for preprocessors, postprocessors, and model.runner.

   • Done when: you know whether specs use type (registry short name) or class_path.

2. Prefer type for built-ins registered in component_factory (e.g. normalize/denormalize patterns in docs).

3. Use class_path + init_args for explicit classes:

   preprocessors:
     - class_path: physicalai.inference.preprocessors.StatsNormalizer
       init_args:
         artifact: stats.safetensors
   

   • Done when: init_args paths resolve relative to the export directory via resolve_artifact.

4. Add a new built-in processor:

   • Implement subclass of Preprocessor / Postprocessor in the appropriate package.
   • Register a short type name in component_factory if manifest-friendly aliases are needed.
   • Add unit tests under tests/unit/inference/preprocessors/ or postprocessors/.
   • Done when: manifest using type or class_path instantiates in a minimal InferenceModel test.

5. Nested components in init_args must stay within _MAX_COMPONENT_DEPTH; avoid cyclic specs.

Validation loop

uv run pytest tests/unit/inference/preprocessors tests/unit/inference/postprocessors tests/unit/inference/test_manifest.py -q

Required checks

• Processor order matches training/export semantics (normalization before runner, denormalization after).
• Artifact file names in manifests do not traverse paths (.., absolute paths).
• New public processors appear in docs/reference/inference-api.md or how-to docs when user-visible.
• Runner choice (SinglePass, chunking runners) stays consistent with predict_action_chunk vs select_action docs.

References

• docs/reference/manifest-schema.md
• docs/how-to/inference/use-manifest.md