Robot learns the concept of direction through motion activity

The concept of direction is one of most important capabilities of a robot in order to better understand the world and accomplish various tasks in the real environment. This paper addresses the problem on how does a robot achieve such a capability. Instead of employing information derived from vision or audition, motion cues are investigated in this study. Specifically, achieving the concept of direction is defined within the robot arm motion domain and the task is twofolds: understanding the arm motion direction (the forward model) and moving arm to some given direction (the inverse model). It is assumed that the robot does not have any dynamic prior knowledge just like human infants do, and the only way the robot can employ is iterative sensory-motor action, so that the concept of direction is established under a developmental style. We equip the robot with an inner learning framework by employing a multimodal deep auto-encoder network, to developmentally learn the concept of direction from each successful trial among its seemingly random sensory-motor actions. One major characteristic of such a framework lies in that the two aspects of the task in learning the concept of direction, i.e. forward model and inverse model, are established simultaneously. In accordance with such an approach, a PKU-HR5II robot successfully achieves the concept of direction, and is able to tell the direction of arm movements with the accuracy of 92.15% and execute arm movements towards given directions with the accuracy of 99.99%.