Pi0 visual normalization mode assumes images to be in the range [0,1]

[atharva-18]

System Info

lerobot version: latest main
OS version: Ubuntu 22.04
Torch version: 2.6.0

Information

• One of the scripts in the examples/ folder of LeRobot
• My own task or dataset (give details below)

Reproduction

policy = Pi0Policy.from_pretrained("lerobot/pi0")
action = policy.select_action(batch)

Expected behavior

Thanks for implementing Pi0! I gave it a shot and observed good results using my robot arm.

Upon further inspection of your code, I found out that the default normalization mode for input images is identity:

lerobot/lerobot/common/policies/pi0/configuration_pi0.py

Line 35 in ee5525f

"VISUAL": NormalizationMode.IDENTITY,

And the policy wrapper makes an assumption that given an identity transform, the input images are already in the range [0, 1], see:

lerobot/lerobot/common/policies/pi0/modeling_pi0.py

Line 360 in ee5525f

# Normalize from range [0,1] to [-1,1] as expacted by siglip

This looks like an important detail which was overlooked. I believe that the default visual transform should be MIN_MAX to avoid incorrect scaling of images to SigLIP or maybe another way is to add an assertion before we prepare images for the backbone.