Are You Using Pre-Trained Policies for Your Research?

I don't think people take pre-trained policies and add stuff on top of that. I could see that maybe for behavioral priors, maybe you don't need to train a policy on a lot of data. If the models are not there, it's very difficult to do this research. Maybe we actually should be releasing our policies now that it's more expensive to train them. It usually takes at least one day or two days depending on the task if you want to move from pixel-based agents to state-based ones.