Unbiased Model-Agnostic Metalearning Algorithm for Learning Target-Driven Visual Navigation Policy

Figure 4

The learning curves of our initial MAML model Ours(loc)/Ours(glo) and UMAML model Ours(UM-loc)/Ours(UM-glo) applied into untrained bedroom scenes. The results demonstrate that Ours(UM-loc) and Ours(UM-glo) all achieve better performances than Ours(loc) and Ours(glo), as the UMAML initial models explicitly minimize the inequality of losses over sampled tasks.