Probabilistic Methods for Imitation Learning in Social HRI

Imitation Learning, also known as Learning from Demonstration (LfD), is a field of study dedicated to aiding an agent's learning process by providing access to expert demonstrations. Within LfD, Movement Primitives is a particular family of algorithms that have been widely studied and implemented in complex robot scenarios. Interaction Primitives, a subset of Movement Primitives, have demonstrated notable success in learning single interactions between humans and robots. However, literature addressing the extension of these algorithms to support multiple variations of an interaction is limited. This thesis presents a physical human-robot interaction algorithm capable of predicting appropriate robot responses in complex interactions that involve a superposition of multiple interactions. The proposed algorithm, Blending Bayesian Interaction Primitives (B-BIP), achieves responsive motions in complex hugging scenarios and can reciprocate and adapt to the motion and timing of a hug. B-BIP generalizes prior work, where the original formulation reduces to the particular case of a single interaction. The performance of B-BIP is evaluated through an extensive user study and empirical experiments. The proposed algorithm yields significantly better quantitative prediction error and more favorable participant responses concerning accuracy, responsiveness, and timing compared to existing state-of-the-art methods.