Master Thesis

Using Symmetries in Reinforcement Learning of Bimanual Robotic Tasks

Work default illustration


  • Started: 10/01/2018
  • Finished: 27/06/2018


Nowadays, the number and variety of robots used in everyday life and production are rapidly increasing.
Most common applications are relegated in the industrial world, where robots stand in the assembly line, executing repetitive works, requiring a high degree of precision, with the aim of increasing the production rate and reducing human presence in dangerous environments. Outside of the factory, robotic applications
in the market are currently able to handle only simple tasks with a low degree of interaction with human beings. One of the new challenge in robotics is to design autonomous robots able to share space with us without the need to modify the infrastructure of our home environment, in contrast to standard industrial
robots, thought to work in predefined factory settings forbidden to humans. This vision creates a whole set of new collaborative and interaction opportunities, but these new environments are also exposed to various sources of perturbation and unpredictable situations not considered before, to which the robot
needs to adapt. In Robot Learning, machine learning methods are used to automatically extract relevant information from data to solve robotic tasks. Modern machine learning techniques, with their power and flexibility, can help to further automate robotics and to narrow the gap towards fully autonomous robots, e.g., for general assistance in households, elderly care, and public services. It becomes crucial in this context to provide these robots with the capacity of generalizing movements and skills, since the range of possible tasks that they could carry out while cooperating with humans is in principle infinite.
One of the key requirements is to make available user-friendly ways of programming robots to "teach" new skills and adapt existing ones to new situations. This possibility is described by Learning from Demonstration [1] paradigm, that provides users with a way to teach robots new knowledge by simply showing some demonstrations, generally by moving the robot himself as it should do to correctly execute the task (kinesthetic teaching). In future perspective, this kind of approach gives to users an instrument to adapt the robot to their needs, regardless of their computer programming skills or experience in robotics.