training configuration for pi0.5

[muxinwu]

Hi, thanks for releasing LIBERO-X.

I'm trying to reproduce the π0.5 results reported in the paper. Could you please share the detailed training configuration used for π0.5 on LIBERO-X (e.g., the training config/yaml or key hyperparameters)?

Thanks a lot!

[muxinwu]

I am trying to reproduce the results using the LeRobot-provided π0.5 base model.

During training, I observed that the training loss quickly decreases and becomes almost stable around 0.5 after roughly 10k steps. Since the loss appeared to converge early, I did not complete the full 30k-step training run and instead evaluated the 25k-step checkpoint.

The paper reports the following results, while my reproduction results are:

Level	Paper	Reproduced	Difference
L1	65.2	56.3	-8.9
L2	53.2	48.4	-4.8
L3	36.0	37.3	+1.3
L4	24.1	31.8	+7.7
L5	18.0	25.7	+7.7

I would like to ask:

1. What are the exact training configurations used to obtain the reported results (learning rate, optimizer settings, random seed, etc.)?

2. Is it expected that the training loss plateaus around 0.5 after about 10k steps when using the π0.5 base model?

3. Based on the numbers above, do my results look reasonable, or do they indicate that there may be an issue in my training/evaluation pipeline?

Could you also share the exact training command or config file used for the reported results?

Any suggestions would be greatly appreciated. Thank you!