Goal state provided as input to model

[KevinyWu]

rank2reward/reward_extraction/models.py

Lines 301 to 309 in c9a40a1

	if not self.state_only:
	# film layer
	if self.film_layer_goal:
	gammabeta = self.film_layer(goal)
	gamma, beta = torch.split(gammabeta, self.r3m_embedding_dim, dim=1)
	x = x * gamma + beta
	else:
	# mix and run through head
	x = torch.cat([x, goal], dim=1)

Hello,

Thanks for sharing the code for your paper! I was wondering why the goal state is passed into the model. In some tasks with the camera angle you use, like drawer-open and drawer-close, the entire arm goes out of the image frame. So wouldn't this goal state help the robot succeed?

Thanks!