Hi, thanks for sharing this work! I had a question about the LIBERO evaluation/training setup and wanted to clarify a few details.
I am trying to understand how the LIBERO experiments were performed, especially in relation to the policy and image generator components.
From the paper/repo, I was unsure about the following:
-
For the LIBERO results, did you use pi0.5_libero or pi0.5_base as the starting policy checkpoint?
-
Was the policy further fine-tuned on LIBERO (Spatial, Object, Goal, LIBERO 10) before evaluation, or was it evaluated zero-shot after training on your dataset?
-
Was imGen further pretrained or fine-tuned on LIBERO data, or was it only trained on the ForeAct dataset?
Thanks in advance!
Hi, thanks for sharing this work! I had a question about the LIBERO evaluation/training setup and wanted to clarify a few details.
I am trying to understand how the LIBERO experiments were performed, especially in relation to the policy and image generator components.
From the paper/repo, I was unsure about the following:
For the LIBERO results, did you use pi0.5_libero or pi0.5_base as the starting policy checkpoint?
Was the policy further fine-tuned on LIBERO (Spatial, Object, Goal, LIBERO 10) before evaluation, or was it evaluated zero-shot after training on your dataset?
Was imGen further pretrained or fine-tuned on LIBERO data, or was it only trained on the ForeAct dataset?
Thanks in advance!