Non-Physical Intervention in Robot Learning Based on LfE Method

This paper proposes an e cient method of robot learning by which a set of pairs of a state and an action are constructed to achieve a goal. Basic ideas of our method are as follows: i) Since autonomous construction of state and action spaces is generally a very di cult problem, we construct a state space so that a group of situations in which an action command to achieve the goal is the same can be merged into one state even if these situations appear to be di erent from each other. An action is de ned as a sequence of the same action command in such a state. ii) Following the LEM (Learning from Easy Missions) paradigm (Asada et al., 1995), we rst nd a set of states (in terms of action) closest to the goal state, and then nd a set of states closest to the set found previously. iii) In order to reduce an enormous number of trials to nd such states, we place a robot so that it can observe objects which the state space consists of (in our case, a ball and a goal). iv) During the above process, the optimal action to achieve the goal is found in every state. This means that a robot can take an adequate action to achieve the goal from every state. We show the experimental results using a real robot system.