[Question]: About goal encoder in InternVLA-N1

[Ideal-ljl]

Question

In the paper, you mentioned "The System 1 model is trained with three major objectives: embedding
alignment among different goals, noise prediction for the diffusion policy, and critic prediction."
But I don't find the training details in the code, is it not prepared?

[wzcai99]

Hi, the entire loss function is implemented here for system-1.

InternNav/internnav/trainer/navdp_trainer.py

Line 100 in b654fbb

loss = 0.8 * action_loss + 0.2 * critic_loss + 0.5 * aux_loss

And the embedding alignment is related to the code here:

InternNav/internnav/model/basemodel/navdp/navdp_policy.py

Lines 206 to 207 in b654fbb

	imagegoal_aux_pred = self.image_aux_head(imagegoal_embed[:, 0])
	pixelgoal_aux_pred = self.pixel_aux_head(pixelgoal_embed[:, 0])