Paul Neculoiu - Migrating robot control systems, towards the universality of robotic brains

Currently robot control systems are specifically designed, engineered and fine tuned for particular problems on particular robots. This leads to a significant waste of man-hours of engineer and Phd level work to implement and reimplement or adapt controllers for similar tasks on different robots resulting in an inefficient robotics industry as a whole. Thus the need to automate or at least semi-automate controller reusability arises. In this project we investigate the hurdles that need to be overcome in attaining controller universality and look into possible methods to bootstrap controllers to the different robot sensors and actuators. A case study was conducted on performing the migration of a controller from a wheeled robot with no mobile vision system (The Pioneer robot) to a legged robot with a mobile head mounted camera (An Aldebaran Nao). The two robots' different modalities makes the task challenging. What does it mean for two different robots to perform the s ame task? Machine learning methods were deployed using artificial neural networks (ANN) to learn the entire sensor abstraction - decision system - robot motor API tree, leaving just sensor feature extraction and low level motor controls in the hands of engineers. The method works reasonably well, effectively linking a number of controllers designed for a Pioneer onto the Nao's sensors and actuators. While preliminary, these methods provide insight into the future prospects of robots programming themselves and learning from each other with the help of humans.