Teaching a robot to play beer pong

Abstract: Contextual Probabilistic Movement Primitives (Contextual ProMPs) extend ProMPs by adding context variables — i.e., variables that do not change dur- ing trajectory execution — to the state representation of a system. In this paper, we use Contextual ProMPs in the beer pong task to generalize demonstrated throwing movements to new locations of the cup. Furthermore, we compare different encod- ings of the context variables, i.e., position of the cup. We approximate the context by a basis expansion in the weight-space of ProMPs using classical conditioning of Contextual ProMPs. In addition, we put the context directly into the state vector using Contextual Linear Regression (CLR), which is equivalent to conditioning of Contextual ProMPs. Using an approximated context together with conditioning of Contextual ProMPs, we achieved a success rate of 70% of hits and 20% of nearly hits. In contrast, CLR was not that convenient in generating such successful throwing movements, achieving only 2 successful and 3 nearly successful hits out of overall 10 attempts using the same set of demonstrations.