Incremental learning from virtual demonstrations and task composition for robotic manipulation

Developing robotic manipulation capabilities that can be incrementally learned, composed, and transferred across platforms remains a significant challenge. In this work, we propose a modular framework that combines imitation learning from virtual demonstrations with reinforcement learning to incrementally train reusable manipulation primitives, which can be flexibly composed into structured tasks. These policies are learned in simplified virtual environments with a limited number of demonstrations and minimal assumptions about the robotic platform or object properties, promoting generality and cross-platform applicability. The learned primitives are also symbolically represented, enabling their reuse and composition through a Hierarchical Task Network (HTN) planning framework. We validate the framework in scenarios involving structured task generation and execution, combining learned and predefined primitives with realistic manipulators and objects. Experimental results demonstrate high success rates in both primitive and long-horizon tasks as well as effective policy transfer across different simulation environments. By integrating incremental skill acquisition with symbolic task composition, our approach provides a modular and scalable solution for adaptive robotic manipulation.