Meta-World Reproduction Issue

[jis3613]

Hi! Thanks for the great work!

I'm trying to reproduce the results of Metaworld for both ML-45 and Fewshot tasks.

I get an average of 84.05% for ML45 and an average of 40.72% for fewshot transfer on held-out tasks across 5 random seeds.

Is there different set of hyper-parameters for Metaworld benchmark?

I'm currently using metaworld_ml45_prise.yaml and metaworld_ml45_prise_fewshot.yaml for each benchmark, respectively.

I'm using quest.yaml which was used in LIBERO for Metaworld as well.

[zhaoyanpeng]

i got 83.0% and 46.4% from a single run. have you found a clue as to why the numbers are so low?

[jis3613]

Not yet

[jayaramreddy10]

I am also observing the similar figures for metaworld, (84.3 for multi task, 38.96 for few shot transfer). Any improvements from your previous runs?