Skip to content

reproducing pi0.5 numbers #103

Description

@quincy-u

Hi thanks for the great work! I was able to run lingbot-VA model locally and was able to get the same number as reported in paper. But I failed to reproduce the number for pi0.5 as reported in paper. Specifically, I run the robotwin data collection locally and regenerate the 2500 clean + 25000 random data and trained pi0.5 using the locally-collected mixed data. The config for pi0.5 follows #12 and I was using jax version of pi0.5. But the success rate I got is around 50-60%.
Easy: 56.86%; Hard: 54.14%.

Could you help me confirm the config for pi0.5? I am also wondering what is your data augmentation used for robotwin? Could the data augmentation contribute to the 20% SR gap? Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions