why use the output of the first decoder layers in ACT model？

[junhui1997]

hs = self.transformer(src, None, self.query_embed.weight, pos, latent_input, proprio_input, self.additional_pos_embed.weight)[0], In the ACT model， should this index be -1？