Pytorch BC/RL Implementations for BridgeData V2

[Stardust-hyx]

Distributed training code (based on deepspeed) for BridgeData is now available!
See https://github.com/Stardust-hyx/Bridge-Torch
Since most open-source LLMs are implemented with Pytorch and require memory-efficient distributed training frameworks to fine-tune on, we believe our repository can greatly facilitate the application of Multimodal LLMs on this field.

We provide implementations for the following methods:

• Goal-conditioned BC
• Goal-condtioned IQL (TODO)

By the way, a novel generative VLM-based BC method is under development by us. Stay tuned.

[Stardust-hyx]

A checkpoint for GCBC as well as a torch-version of eval.py is also provided in https://github.com/Stardust-hyx/Bridge-Torch. We will be grateful if you can help check the validity of the checkpoint by trying it out. @HomerW
Thanks in advance!