Unsupervised approach to acquire robot joint attention

Existing approaches to joint attention of robots have considered an object's information (location and visual information) with a head pose of caregiver using data from many subjects to train the robot joint attention model. These approaches have used simulated data (of the object) to train the robot joint attention model, and they are incapable of accurately predicting a caregiver's attention when the caregiver has a complex eye gaze pattern. A complex eye gaze pattern can be defined as a caregiver going over the number of objects in the environment and finally attending to the object of interest. At this time it is possible for us to obtain a long sequence of eye gaze data using a combination of different eye gaze patterns. Our approach segments the eye gaze data and applies a Mixture Gaussian-based unsupervised cluster to detect the caregiver's intention at each of the time segmentations. Finally, the above attention information is combined with a geometrical model of objects to detect the caregiver's object of interest by considering the entire eye gaze segmentations. The novelty of our approach is to detect the caregiver's object of interest when the caregiver has a complex eye gaze pattern and does not even use any of the training data. The experimental results revealed that when the objects distance is 20cm, our proposed approach can accurately recognize and impressive 80% of the caregiver's interested objects. The contrivance of the time segmentation is manipulated to infer a caregiver's attention plans and behaviors in each time interval. It is directed to detect the caregiver interested object for acquiring the skills of joint attention.