inference details, adapt base Pi0.5 with RECAP
Hello, so I know you trained on the lerobot aloha_sim_transfer_cube_human dataset. Did you run any sort of inference after training in simulation or real hardware to validate how well the model performs?
I would suspect training on such a small dataset with ~6 hours on 8x A100 80GB you would be in danger of overfitting right?
To circumvent this, I am wondering if there is a way to modify the Pi0.5 repository already released to incorporate RECAP. In other words: keep the Pi0.5 base model weights (which already leverages 10k+ hours of real robot teleoperation data which we cannot reproduce), but add both RECAP and fine tuning on top of it.
This way, we may get the benefits of both the large dataset and RECAP.
I'm imagining something like this:
1.) Load Pi0.5 base model weights (official repository).
2.) Freeze early layers of the VLM (or set very low LR).
3.) Add RECAP stuff
4.) Policy warmup: Run behavior cloning on lerobot aloha_sim_transfer_cube_human.
5.) RECAP training: Apply advantage-conditioned updates only on the value head, keeping the base VLM mostly frozen.
6.) Evaluate in sim or real aloha hardware
Adding recap to a base model may be "too invasive" where complete retraining of the model is required... not sure.
For those interested in the above post, please see this: