feat: MolmoAct2 processors

[alfieroddan]

Summary

Corresponding PR to introduce numpy processors

open-edge-platform/physical-ai-studio#809

Other changes: we've adjusted the stats normalization to include the new mask parameter.

[Copilot]

Pull request overview

This PR adds NumPy-based MolmoAct2 inference processors to the runtime (preprocess: observation → prompt/images; model-input assembly: tokens/images → graph inputs; postprocess: outputs → actions), and extends stats (de)normalization to support a per-dimension boolean mask so only selected dimensions are transformed.

Changes:

• Introduce MolmoAct2 preprocessor, image patchifier, model-input assembler, and postprocessor, and register them in the component factory.
• Add masked-dimension support to StatsNormalizer / StatsDenormalizer and corresponding unit tests.
• Add new unit tests for MolmoAct2 preprocessing/model-input assembly/postprocessing pipeline.

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
tests/unit/inference/preprocessors/test_stats_normalizer.py	Adds a masked-dimension normalization test for StatsNormalizer.
tests/unit/inference/preprocessors/test_molmoact2.py	New unit tests for MolmoAct2 preprocessor/image processor/model-input assembly + manifest pipeline.
tests/unit/inference/postprocessors/test_stats_denormalizer.py	Adds a masked-dimension denormalization test for StatsDenormalizer.
tests/unit/inference/postprocessors/test_molmoact2.py	New unit tests for MolmoAct2 postprocessor and registry instantiation.
src/physicalai/inference/preprocessors/stats_normalizer.py	Applies optional mask after normalization so unmasked dims pass through unchanged.
src/physicalai/inference/preprocessors/molmoact2.py	New MolmoAct2 observation preprocessor (task normalization, state discretization, image packing/resizing, optional joint transform, optional masked normalization).
src/physicalai/inference/preprocessors/molmoact2_inputs.py	New MolmoAct2 model-input assembler: expands `<
src/physicalai/inference/preprocessors/molmoact2_image.py	New NumPy image patchification + pooling index generation for MolmoAct2.
src/physicalai/inference/preprocessors/init.py	Exposes MolmoAct2 preprocessor and model-input assembler.
src/physicalai/inference/postprocessors/stats_denormalizer.py	Applies optional mask after denormalization so unmasked dims pass through unchanged.
src/physicalai/inference/postprocessors/molmoact2.py	New MolmoAct2 postprocessor (clamp, masked denormalize, optional inverse joint transform).
src/physicalai/inference/postprocessors/init.py	Exposes MolmoAct2 postprocessor.
src/physicalai/inference/component_factory.py	Registers MolmoAct2 processors in the component registry.
pyproject.toml	Minor dependency-list whitespace change.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[kdijkstraintel]

Looks good.

[samet-akcay]

would it be an idea to re-structure src/physicalai/inference/postprocessors directory as it is getting a bit cluttered? Maybe a subpackage per-model? Something like

  src/physicalai/inference/preprocessors/
  ├── __init__.py
  ├── base.py
  ├── stats_normalizer.py
  ├── resize.py
  ├── pi05.py
  ├── rldx1/
  │   ├── __init__.py
  │   ├── processor.py
  │   ├── rope.py
  │   └── token_composer.py
  └── molmoact2/
      ├── __init__.py
      ├── processor.py
      ├── inputs.py
      └── image.py

init.py could expose the classes:

  from physicalai.inference.preprocessors.molmoact2.inputs import MolmoAct2Inputs
  from physicalai.inference.preprocessors.molmoact2.processor import MolmoAct2Preprocessor

  __all__ = ["MolmoAct2Inputs", "MolmoAct2Preprocessor"]

so relative imports would still work?