[Fix] Fix the path resolve issue in windows

[whats2000]

Fix: Fix the path resolve issue in Windows

What this does

This PR fixes a path-resolution issue on Windows by ensuring that pretrained_path is kept as a string rather than converted to a Path object.
Using Path() caused backslash/forward-slash inconsistencies, especially when handling HuggingFace repo IDs.
The fix ensures consistent forward-slash formatting across platforms.

How it was tested

• Updated eval.py and train.py to avoid using Path() for pretrained_path where forward-slash preservation is required.
• Verified that Windows paths resolve correctly when loading configs.
• Ran local training/eval runs on Windows to ensure no regressions during config loading.

How to checkout & try? (for the reviewer)

To test the behavior locally run a training command that loads a pretrained policy:

lerobot-train --config-path=path/to/config.yaml

[Copilot]

Pull Request Overview

This PR fixes a path resolution issue on Windows by preventing the conversion of pretrained_path to a Path object, keeping it as a string instead. This ensures consistent forward-slash formatting, which is essential for HuggingFace repository IDs and cross-platform compatibility.

Key changes:

• Modified path handling in train.py and eval.py to preserve pretrained_path as a string
• Used as_posix() method when converting Path objects to strings to maintain forward slashes

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/lerobot/configs/train.py	Updated to keep pretrained_path as string and use as_posix() for path consistency
src/lerobot/configs/eval.py	Updated to keep pretrained_path as string instead of converting to Path object

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[whats2000]

This kinda because the path in the window will use the \\ for path; if we use the path object, this will always make the HuggingFace repo invalid to resolve. So a better way is that we keep them in posix style string to ensure all the paths create by train and resolve by train in posix style!

[DanielBryars]

Ahhh I ran into this too. #2552