[Bug] Diffusion Policy: Missing image resize logic for high-res datasets (e.g., aloha_sim)

[fxdllr]

Ticket Type

🐛 Bug Report (Something isn't working)

Environment & System Info

LeRobot : 0.4.3
Python : 3.13.11

Description

Summary:
I'm training Diffusion Policy using the official aloha_sim_transfer_cube_scripted dataset (640x480 resolution).
The default DiffusionConfig sets crop_shape=(84, 84), but there is no automatic resize step in the pipeline.
The Problem:
The validation logic in configuration_diffusion.py only checks if crop_shape < image_shape. Since 84 < 480, it passes.
However, during training, RandomCrop(84) is applied directly to the 640x480 image. This means the model only sees a tiny crop (2.7% of the image), often missing the robot arm or object entirely.
Expected Behavior:

1. The code should warn the user if image_shape is significantly larger than crop_shape.
2. Or, ImageCropResizeProcessorStep should be automatically configured to resize images to something close to crop_shape (e.g., 96x96) before cropping.
   Impact:
   Users training on official datasets get very poor success rates because the model is training on empty crops.

Context & Reproduction

No response

Relevant logs or stack trace

Checklist

• I have searched existing tickets to ensure this isn't a duplicate.
• I am using the latest version of the main branch.
• I have verified this is not an environment-specific problem.

Additional Info / Workarounds

No response