Novel robotic skill synthesis with Conditional Neural Movement Primitives

Humans are capable of executing a wide variety of complex tasks, based on prior experience. Often, they accomplish them by breaking them down into minor actions that are composed together one after the other to achieve the goal. These actions are not always learned directly but adapted from previous similar experiences to the current context. In this study, we propose a computational model that is biologically inspired and aims to integrate into robotics the human ability to adapt movements and combine them to achieve high-level skills. A novel approach to high-level skill synthesis is explored by leveraging movement primitives learned through Conditional Neural Motion Planning (CNMP) models. The research introduces two methods for generating and composing new actions based on demonstrated ones. In the first approach, trajectories are blended by utilizing the task interpolation capabilities of the neural network and a developed mathematical system for parameterization. Additionally, two alternative architectures for the CNMP model are proposed, both achieving results comparable to the original model while accommodating partial information. The second approach achieves action synthesis through the concatenation of primitives, spatial interpolation, and the network's ability to encode multidimensional data to embed the environment representation. Both proposed methods are finally showcased in several experiments with real robots.