Why use agent-centric data then use wrapper to support scene-centric in roll-out?

[daeheepark]

During roll out, it uses EnvUnifiedSimulation to get observation in script/scene_editor.py

In line 259 of EnvUnifiedBuilder, it is written that "# hardcode to agent-centric and then use a wrapper to support scene-centric". However, what is the reason behind that?

Actually, I tried to get observation from UnifiedDataset with scene-centric manner, and converted them to agent-centric with "convert_scene_data_to_agent_coordinates" method from tbsim.utils.trajdata_utils.py. Then, the SceneDiffuserTrafficModel works, but the results look weird. (It only predicts straight trajectories)

What is the proper way to roll-out trajectories with scene-centric data from UnifiedDataset?

[daeheepark]

I have wrote the predict_step function in SceneDiffuserTrafficModel as above. The output action only contains straight trajectories as below: