[YeDonS]

Both SO101 CLI entrypoints pass a non-empty --calibration-dir directly from argparse (a string) into SO101FollowerConfig. The config's Path | None annotation does not coerce it. SOFollower.__init__ then calls Robot.__init__, which executes:

self.calibration_dir.mkdir(parents=True, exist_ok=True)

As a result, supplying a custom calibration directory raises AttributeError: 'str' object has no attribute 'mkdir' before robot connection. The inference entrypoint reaches this after policy initialization.

Convert the value to Path at both configuration boundaries. This makes custom calibration directories usable for calibration and inference. Both files already import Path. Keeping the existing conditional preserves omitted arguments and the explicit empty-string fallback; relative paths retain their existing interpretation.

Validation

On macOS arm64, an isolated harness runs both complete CLI modules with the repository's actual config classes, SOFollower/Robot constructors, MotorCalibration, and calibration-file load/save methods. Python 3.10.20 and 3.12.13 with draccus==0.8.0 each verify 28 scenarios:

Input category	Before	After
Omitted / empty directory (4 cases)	Succeed	Succeed
Valid non-empty paths / calibration-file operations (18 cases)	mkdir AttributeError	Succeed
Invalid directory / malformed JSON (6 cases)	mkdir AttributeError masks the actual problem	Expected filesystem / JSON errors

Coverage includes relative and absolute paths, spaces, Unicode, existing directories, existing calibration reads without modifying their bytes during initialization, save/reload roundtrips, and custom robot IDs. Compilation and git diff --check pass.

Hardware methods/factories and model/tensor dependencies are stand-ins; the runtime policy loader is replaced and its control loop uses zero steps. This validates path and calibration-file handling; real robot calibration, M50 execution and model performance were not tested.

Reproduction harness, commands and before/after reports: isolated verification bundle.