get_expert_actions() returns all-zero actions when using Waymo dataset — imitation learning impossible

[duguxiaohun]

Hi, thanks for your great work on GPUDrive!

I found a potential issue in the function get_expert_actions().
According to the docstring, this function should return:

expert_actions: Inferred or logged actions for the agents.

However, when using the Waymo dataset:

The dataset does not contain logged actions;

The current implementation returns an all-zero action tensor (e.g., shape [100, 64, 3]);

As a result, any downstream module that relies on expert actions (e.g., imitation learning, behavior cloning, IRL) receives meaningless actions.