evgeniishch
commented
5 months ago return { 'uc_actions': uc_actions, 'gc_actions': gc_actions, 'gc_distance': gc_distance, }
It corresponds to "goal-conditioned" and "unconditioned" modes described in the paper. The difference is that in unconditioned mode the goal token is masked so the model generates trajectories for exploration.
return { 'uc_actions': uc_actions, 'gc_actions': gc_actions, 'gc_distance': gc_distance, }