openvinotoolkit/physicalai/pai-configure-inference

Configuring the Inference Pipeline

Pipeline order: observation → preprocessors → runner → postprocessors → action output. See the pre/post-processing guide and manifest schema.

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.

1. Prefer `type` for built-ins registered by your installed Runtime package (e.g. normalize/denormalize patterns in the docs).
2. Use `class_path` + `init_args` for explicit classes:

```yaml 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.

1. Validate the manifest with a representative export by constructing InferenceModel and running a fake observation; verify preprocessing precedes the runner and postprocessing follows it.

• Done when: the model returns an action with the expected shape.

1. Nested components in init_args must stay within Runtime's configured depth limit; avoid cyclic specs.

Validation loop

Run a one-observation smoke inference in the user's project after each manifest change; do not connect robot hardware to test a config edit.

Required checks

• Processor order matches training/export semantics (normalization before runner, denormalization after).
• Artifact file names in manifests do not traverse paths (.., absolute paths).
• A named type exists in the installed package; use class_path only for trusted local components.
• Runner is single_pass (SinglePass), the only built-in runner; do not invent other runner types. Its use stays consistent with the predict_action_chunk vs select_action docs.

References

• Manifest schema
• Using a manifest