we recently have met some problems with the data loading function: is there a way that we can get the state and action from the demonstration data? for example, in the format of state-action (74-dimension state features + 2-dimension actions ) pairs (just like those defined in the controlling environment), but the action follows the expert trajectories in the highD or inD dataset. In this way, some inverse-RL algorithms can be built upon your package.