Physicalai Runtime Working With Config

Working with physicalai.config

Runtime owns src/physicalai/config/. Physical AI Studio should import this package from physicalai, not copy it under library/.

Workflow

1. Pick the API
   • Portable YAML recipes (robots, cameras, exported components): Config and @export_config.
   • Known Python types (trainers, dataclass configs, CLI models): jsonargparse (ArgumentParser, add_class_arguments, parse_object, instantiate).
2. Author a recipe — class_path + init_args; nest recipes only for trusted local config. See docs/how-to/config/instantiate-components.md.
   • Done when: dict/YAML passes validation without ConfigError.
3. Export live objects — @export_config, then Config.from_instance(obj) and Config.save(). Only trusted local sources.
   • Done when: saved YAML reloads with instantiate() in tests.
4. Typed construction — use jsonargparse in the owning package (runtime CLI, inference, and so on). Avoid new generic loaders under physicalai.config.
5. Document and test — update docs/explanation/configuration.md or docs/how-to/config/; extend tests/unit/config/.

Validation loop

uv run pytest tests/unit/config/ -q
prek run ruff-check --all-files

Required checks

• Deeply nested config hits _MAX_CONFIG_DEPTH and raises ConfigError.
• class_path comes only from trusted local config (docs/development/security.md).
• Studio imports Config from runtime, not a duplicate tree in the library.

References

• docs/explanation/configuration.md
• docs/how-to/config/instantiate-components.md
• tests/unit/config/