Training configuration for the π0.5 RoboTwin 2.0 baseline

[QiuGuangwww]

Hi, thanks for the great work!

I am trying to reproduce the π0.5 results reported on RoboTwin 2.0 (82.74% on Clean and 76.76% on Randomized scenes).

The paper specifies that π0.5 was fine-tuned from a pretrained checkpoint using 2,500 Clean demonstrations and 25,000 Randomized demonstrations, but I could not find the detailed training hyperparameters for this baseline.

Could you please share the exact training configuration used for π0.5 on RoboTwin 2.0, especially:

• global batch size
• number of training steps or epochs
• learning rate and learning-rate schedule
• optimizer and weight decay
• warmup steps
• number/type of GPUs
• the exact pretrained π0.5 checkpoint
• whether the Clean and Randomized results were obtained from separate training runs or a single jointly trained model

If possible, sharing the corresponding OpenPI config or training command would be extremely helpful for reproduction.

Thank you!