From demonstration to execution: Single-demonstration task and motion learning for robotic manipulation
Year: 2026
Type: article
Abstract: Learning from demonstration (LfD) provides an intuitive approach for enabling intelligent systems to acquire manipulation skills; however, most existing methods either require multiple demonstrations or treat task reasoning and motion generation separately, limiting generalization in complex scenarios. This paper proposes a unified LfD-based task and motion planning framework that learns executable task-motion knowledge from a single demonstration and generalizes it to similar tasks. An agent-centric action segmentation module is introduced to extract execution-oriented action primitives from the demonstration, which are encoded using semantic network-based representations for structured task knowledge modeling. Spatial relations are extracted and transformed into PDDLStream operators, enabling structured task representation and reasoning with integrated geometric constraints. For motion generation, an artificial potential field-based dynamic movement primitive is adopted to produce adaptive and collision-free trajectories. By tightly coupling symbolic task reasoning and continuous motion generation, the proposed framework enables reproduction and generalization of demonstrated tasks under varying object configurations and environmental constraints. Experimental evaluations on stacking and assembly tasks, together with comparisons against four state-of-the-art methods over 30 repeated trials, demonstrate the effectiveness and robustness of the proposed framework for complex robotic manipulation.
Source: Advanced Engineering Informatics
Authors: Tao Ma, Mingkun Wu, Burkhard Corves
Institution: RWTH Aachen University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,560
OpenAlex ID: W7233555692
