Human Action Recognition Oriented to Humanoid Robots Action Reproduction

Our research aims at providing a humanoid robot with the ability of observing, learning, and reproducing actions performed by humans in order to acquire new skills. In other words, we want to apply artificial intelligence techniques to automatically recognize a human activity in order to make a humanoid robot able to reproduce it.This system has not only to distinguish between different actions, but also to represent them in a proper manner to allow a robot to reproduce the motion trajectories the demonstrator showed and learn new skills. Since the final system is going to be integrated in an autonomous humanoid robot (specifically model Aldebran Nao), we are working with an RGB-D sensor (Microsoft Kinect) that can be easily applied to it. This objective introduces also strict real-time constrains to the action recognition algorithm: we have opted for a probabilistic approach that offers good modeling and fast recognition performances.