Learning and Generalizing Tasks on Humanoid Robots with an Automatic Multisensory Segmentation Method

We provide a complete framework for learning and reproducing tasks from human demonstrations. This framework adapts recent developments in automatic, unsupervised segmentation of time-series to humanoid robotics by preprocessing the data obtained from a broad range of the robot's sensors, to then repropduce the learned task in similar environments. In more detail, we reproduce and extend the acquired multi-step task using Dynamic Movement Primitives in simulation for the JVRC1 Robot, and further validate it the segmentation process in real world with the HRP-4C Robot, thus showcasing the possibility to create an extensive library of reusable skills for complex humanoids with our approach.