We use the feature-level supervision from the roach expert which is a learned model. And we use its predicted control distribution parameters for control action prediction. You can refer to the Roach paper for more details.
We find that the roach expert performs better than the official rule-based expert.
I am curious about a question, why not use an official expert agent to collect data(auto_pilot.py), but rather use a roach (roach_ap_agent.py)