Machine learning provides powerful tools for solving problems in robotics that, due to their complexity, cannot be addressed solely by engineering. Object recognition is a prime example of a robot vision problem that can hardly be solved without any learning. Grasping and manipulating unknown objects is another type of a robotic task that is difficult to solve by hand-coding. Robotics also provides researchers in machine learning with interesting new challenges. For instance, the field of reinforcement learning has been deeply shaped by problems that typically arise in robotics, such as partial observability. Another example is learning by demonstration, which is a particular type of supervised learning that is specific to robotics.
The research area at the intersection of machine learning and robotics is known as robot learning. The question driving research in this area is: How can a robot learn from experience to perform complex tasks? My talk will focus on problems in machine learning that originate from robotics. In particular, I will present challenging problems related to reinforcement learning and to learning by demonstration. I will show how these problems can be efficiently solved by formulating them as optimization ones, derived from high principles such as entropy minimization. I will also show how the proposed algorithms were applied in robotics.