search

google-deepmind / open_x_embodiment

Apache License 2.0
623 stars 41 forks source link

Can the model be fine-tuned? #21

Closed reginald-mclean closed 7 months ago

reginald-mclean commented 7 months ago

Hi! Thank you so much for providing the model checkpoint and starter code to use the model. Is it possible to fine tune the RT1-X model? We are having some issues applying the model to our environment, but fine-tuning it could fix the issue. Thanks in advance!

open-x-embodiment commented 7 months ago

Hi! It should be possible to finetune the RT-1-X model on your own. Can you please elaborate what issues you are facing? Also, we are hoping to release Jax code with example finetuning code. That should be a nice reference code for that.

Best, Pannag

On Thu, Nov 9, 2023 at 11:29 AM 'Reggie' via Open X-Embodiment Collaboration @.***> wrote:

Hi! Thank you so much for providing the model checkpoint and starter code to use the model. Is it possible to fine tune the RT1-X model? We are having some issues applying the model to our environment, but fine-tuning it could fix the issue. Thanks in advance!

— Reply to this email directly, view it on GitHub https://github.com/google-deepmind/open_x_embodiment/issues/21, or unsubscribe https://github.com/notifications/unsubscribe-auth/BDPTRVPN5PX2JBN7YYNP4QLYDUAHBAVCNFSM6AAAAAA7E27XVKVHI2DSMVQWIX3LMV43ASLTON2WKOZRHE4DKOJZGQ3TANA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

-- You received this message because you are subscribed to the Google Groups "Open X-Embodiment Collaboration" group. To unsubscribe from this group and stop receiving emails from it, send an email to @.*** To view this discussion on the web visit https://groups.google.com/d/msgid/open-x-embodiment/google-deepmind/open_x_embodiment/issues/21%40github.com https://groups.google.com/d/msgid/open-x-embodiment/google-deepmind/open_x_embodiment/issues/21%40github.com?utm_medium=email&utm_source=footer .

reginald-mclean commented 7 months ago

I am attempting to use RT-1-X model in Meta-World. I used the code from the minimal example notebook to load and use the policy. Meta-World has an action space of size 4, the desired xyz position of the end effector and a scalar value that indicates how open/close the gripper should be. I took the base displacement, base displacement vertical, and gripper open/close values from the policy action dictionary and use those as the action (concatenated together). However, I am not seeing any tasks being solved with this method.

quanvuong commented 7 months ago

Hi Reggie,

For the RT-1-X model, we currently do not expect the model to zero-shot solve manipulation problems in a new domain that the model was not trained on (such as Meta World).

The base displacement and base displacement vertical are also intended to control the base movement of a robot, and not the xyz position of the end effector.