no pretrained unimanual checkpoints used during training

[Wuli10]

Hi, thanks for open-sourcing this work.

I am confused about the pretraining part in the paper vs. the current repo.

From the paper, AnyBimanual seems to rely on pretrained unimanual PerAct policies, and the results in Table 1 also includes PerAct2 + Pretraining. Although part of the ckpt was released in the README.md. However, in the public code I could not find a clear path that:

• loads pretrained unimanual checkpoints,
• maps them into the bimanual / AnyBimanual model,

I notice that in the conf/config.yaml, it contains the framework.use_pretrained and the framework.load_existing_weights flags, but they are both set False.

Questions

1. Is framework.use_pretrained currently unused / deprecated?

2. For the reported AnyBimanual results, were pretrained unimanual PerAct checkpoints explicitly loaded?

   • If yes, where is that logic in the repo?
   • If not released yet, could you briefly describe the intended loading / initialization process?

3. For the PerAct2 + Pretraining baseline:

   • What pretrained weights were used? Were they pretrained PerAct checkpoints?
   • If they were PerAct checkpoints, how were they transferred into the PerAct2 architecture? Since the architecture is a bit different between PerAct and PerAct2.

4. The paper mentions pretraining on 18 RLBench unimanual tasks:
   • Are those checkpoints available?
   • Is there a script/config to reproduce that stage?

Right now, I cannot tell whether this part is missing from the public release, or whether I am looking in the wrong place.

Thanks for any clarification.