Configure Preprocessing and Postprocessing

Preprocessors run before model execution. Postprocessors run after model execution. Together, they define the input and output adaptation around the runner.

The manifest can declare both stages.

model:
  runner:
    type: action_chunking
    chunk_size: 50
  artifacts:
    openvino: model.xml
  preprocessors:
    - type: normalize
      artifact: stats.safetensors
  postprocessors:
    - type: denormalize
      artifact: stats.safetensors

The same components can also be declared with explicit class paths.

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

Pipeline shape:

observation
  -> preprocessors
  -> runner
  -> postprocessors
  -> action output

Use type for registered built-in components. Use class_path when you want an explicit import path.