robopen/roboagent

Finetune on new data

Closed this issue · 3 comments

Very good paper, thank you for the release.

How much data do you need to fine tune the model to work on new environment and task?

To fine-tune a pre-trained model, ~25-75 trajectories should be good enough.
We were able to get a new task in a new environment to work with ~25 trajectories fine-tuned

Thank you for the answer. This was with 4 cameras? did you try with 1 camera on real world data, how was the performance?

We ran exps with 2 cameras + new setup/location and the performance loss was less than 10%