nicklashansen / tdmpc2

Code for "TD-MPC2: Scalable, Robust World Models for Continuous Control"
https://www.tdmpc2.com
MIT License
343 stars 71 forks source link

[Q]: Multi-task Meta World Check points #22

Closed Ozzey closed 6 months ago

Ozzey commented 6 months ago

Amazing work with both TD-MPC 1 and 2!

I wanted to know if you have any pretrained stand-alone Meta World multi task mode/checkpoint? I can see there is standalone model which only supports DMControl tasks but not one for Meta World. It is possible to use the 80task DMControl + Meta World model but it is a bit too computational heavy.

If not, do you have any available Meta World dataset for training a standalone multi-task model?

Best regards,

nicklashansen commented 6 months ago

Thank you! We did not train a standalone multi-task model for the Meta-World tasks. However, we have released both single-task checkpoints and the dataset used for training the 80-task checkpoints, so depending on your use case you should be able to use either of them! Iirc the 80-task checkpoint is doing quite well on Meta-World as well, so if you don't want to train anything perhaps you can use one of the smaller checkpoints as a starting point?

Ozzey commented 6 months ago

Thank you so much for the response! Currently I was testing mt80 checkpoint of size 48M since it seemed to be the most balanced one, but still since I wanted to use it only for manipulators (and hence Meta-World task), using mt80 itself is a bit inefficient for my use-case. Which is why I am trying to train and evaluate the model purely based on 50 Meta-World tasks.

So, it would be great if I could get the either the single task data set for Meta-world or the multi-task dataset for the 50 tasks. If not, I still appreciate the help, thanks a lot once again!

nicklashansen commented 6 months ago

Sure! The 80-task dataset is available here: https://www.tdmpc2.com/dataset You can filter tasks by ID; the last 50 task IDs should correspond to Meta-World tasks.

Ozzey commented 6 months ago

Thank you! I'll try it out.