Development of Integrated Behaviour in a Simulated Humanoid Robot Exploiting Language Assisted Training and Self Talk

In this project, a simulated humanoid robot learns to reach for, grasp and move a spherical object to a desired target location. The trial and error training process is divided into multiple stages where the robot first develops lower-level behaviours that it can later integrate to achieve the higher-level goal. The robot is supported by a caretaker who provides linguistic labels indicating the action that should currently be performed. Multiple approaches to teaching the robot to become autonomous (no longer rely on the caretaker) are explored. One such technique is to let the robot engage in a form of self talk, where it can give itself the linguistic instructions used by the caretaker. The robots trained with the proposed method are able to accomplish all the desired behaviours. The results of the experiments suggest that the lower-level skills serve as scaffolding for the integrated behaviour. The self talk mechanism improves the robustness of the autonomous behaviour, and the ability to adapt it to situations that are not experienced during training. Some individuals are even able to adapt better to new situations when they are acting autonomously than when they are guided by the caretaker.