google-deepmind / open_x_embodiment

Apache License 2.0
774 stars 55 forks source link

Data Standardization Preprocessing #12

Open elin-bdai opened 10 months ago

elin-bdai commented 10 months ago

We've been exploring the RT-X dataset and noticed that though all the datasets are the same tfrecord format, they aren't standardized across datasets (understandably so with the many types of robots). There was a little bit of text in your paper about standardization, but not a lot. We were wondering if there are any pre-processing scripts that standardize the datasets before feeding it into RT-1/RT-2 that you could share as an example? Thank you!

kpertsch commented 10 months ago

Hi!

We are working on a code release for an RT-X training pipeline in Jax, which will also include the action standardization and normalization code, please stay tuned! In the meantime, you can check out the RT-X model checkpoint Colab for an example of pre-processing and normalization on the Bridge dataset. It shouldn't be hard to adapt to other datasets too!