based on pi0.5

[FishofJLlake]

Thank you for the great work on open source.

What extra work is needed when training a reinforcement learning model based on pi0.5 compared to pi0?

[akshay18iitg]

Rn, pi05 has elements like response prediction, discrete action, and knowledge insulation that pi0 doesnt have. So, adding an indicator token before discrete action prediction and after response prediction (especially while inferencing) will be most of the work.

Currently, we are supporting RL for pi0. Eventually, we will use it for pi05 as well.

[FishofJLlake]

Rn, pi05 has elements like response prediction, discrete action, and knowledge insulation that pi0 doesnt have. So, adding an indicator token before discrete action prediction and after response prediction (especially while inferencing) will be most of the work.

Currently, we are supporting RL for pi0. Eventually, we will use it for pi05 as well.

Thanks for your reply! I'm really looking forward to your subsequent work.