Predicting Supportive Behaviors for Human-Robot Collaboration

We present a model for predicting what supportive behaviors a robot should offer to a person during a human-robot collaboration (HRC) scenario. We train and test our model in simulation, using noisy data that mimics a real-world HRC interaction. Our results show that we can achieve accurate predictions, using only a small set of labeled demonstrations. We also show transfer learning capability: we train our model on an initial task and test it on a new task composed of the same building blocks but structured differently.