We tested seven octopuses, Octopus vulgaris, in maze-learning experiments. They tried to reach the goal, so as to get a reward, by using various locomotory actions in the path, and sometimes encountered obstacles. They came to select efficient swimming actions in the path; afterwards less efficient tactile actions (crawling, staying put, and so on: these reduce the speed of movement) gradually increased, while time to detour around the obstacle was reduced. To investigate whether octopuses reduce time spent detouring around obstacles by estimating their actions in the path, we devised a trade-off situation in which octopuses were obliged to use tactile actions even though the set-up also encouraged them to use swimming actions. As a result, we could observe that they reduced the detouring time. In that way, we experimentally constituted a perspective as if octopuses looked around the whole maze and estimated their actions. Such a perspective appeared to be autonomous learning.