Task-Adaptive Robot Learning From Demonstration With Gaussian Process Models Under Replication

Learning from Demonstration (LfD) is a paradigm that allows robots to learn complex manipulation tasks can not be easily scripted, but demonstrated by human teacher. One of the challenges LfD enable acquire skills adapted different scenarios. In this letter, we propose achieve exploiting variations in demonstrations retrieve an adaptive and robust policy, using Gaussian Process (GP) models. Adaptability enhanced incorporating task parameters into model, which encode specifications within same task. With our formulation, these either real, integer, or categorical. Furthermore, GP design exploits structure replications, i.e., repeated with identical conditions data. Our method significantly reduces computational cost model fitting tasks, where replications are essential obtain model. We illustrate approach through several experiments on handwritten letter demonstration dataset.