What is the expected MAE for pi0.5 triton implementation?

[kianhk6]

As mentioned in the title, what are your guys claim for the MAE? Or the relative error for the pi0.5 implementation?

Because this is what I currently get and I was wondering if its expected or it can be lowered?

[kianhk6]

This MAE result leads to a low LIBERO evaluation (10% drop in success score) @Tomorin523 would be appreciated if you could confirm that with me

[Tomorin523]

can you provide your test.py and pi05_infer.py? The result is apparently abnormal.