"Teaching robots to do tasks is a more challenging task than it may appear at a first glance, even when these robots resemble us. In this presentation, I will shed some light on some of the major challenges we face when transferring skills from humans to robots.
Imitation learning is not just play and record. A key component of learning is the ability to generalize. In imitation learning, generalization must come from very few examples to ensure that the length of the training phase is bearable to the human teacher. Most approaches, however, discard erroneous training examples, hence forcing the human teacher to be very skilled at the task, and to repeat the tasks numerous times until enough successful demonstrations have been gathered. There is an advantage to include errors in the training to ensure robustness and better generalization. I will hence discuss some promising avenues in learning from failed demonstrations.

Traditional planning approaches seek to find the optimal solution. The strength of human control, however, lies in that we are capable of performing the same tasks in multiple ways, several of which being sub-optimal. Feasibility is hence more important than optimality. It offers the possibility to rapidly switch across control strategies in the face of perturbations. Modeling the variability with which humans perform the same task provides robots with a notion of feasibility regions in sensori-motor space. Over the years, we have sought to provide robots with controllers that allow instantaneous reactions to perturbation, mimicking humans’ immediate response in the presence of danger. I will describe application of such fast and robust adaptation for compliant control during manipulation of fragile objects and for performing sports, such as when playing golf with moving targets and when catching fast flying objects.

"