Reward-Based Learning

This project completed our first full humanoid childbot real-time learning behavior. In this scenario, the childbot projected on a large screen to a student actor is presented with an arbitrary sequence of colored/patterned cards (seen through its camera below the screen). If the childbot points to the ball of the same color, the student “rewards” the childbot by pressing a reward number (e.g. 10) on the keypad. If the childbot points to the ball of the wrong color, no reward is given. The student continues in this way until the childbot seems to have learned the “color/pattern” response. Over the course of many reinforcements, the childbot successfully learned to point to the correct color. (Please see video).

Our near future plan is to substitute emotional speech content for the touch pad reward.