Why the training code snippet contains BEV feature and doesn't have language output ?

[HuangChiEn]

Hi, i have 2 question about the codebase :

1. Why the training code snippet contains BEV feature ? (carllava only use RGB image as input)
   

   ETA/carformer/carformer/ponderer.py

   Line 335 in aa510c1

   labels["input_bev_latent"] = bev

2. Most of VLM sacrifices speed for explainable (language) output enhancing the reasoning & robustness of predicted waypoint. However, i didn't saw any language output in the codebase of ETA ?