One-shot learning-driven autonomous robotic assembly via human-robot symbiotic interaction

Abstract

Multi-procedure robotic assembly requires robots to sequentially assemble components, yet traditional programming is labor-intensive and end-to-end learning methods struggle with vast task spaces. This paper introduces a one-shot learning from demonstration (LfD) approach that leverages third-person visual observations to reduce human intervention and improve adaptability. First, an object-centric representation is proposed to preprocess demonstrations of human assembly tasks via RGB-D camera. Then, a kinetic energy-based changepoint detection algorithm automatically segments procedures, enhancing the robot’s understanding of human intent. Third, a demo-trajectory adaptation-enhanced dynamical movement primitive (DA-DMP) method is proposed to improve the efficiency and generalization of motion skills. The integrated system uses visual feedback for closed-loop reproduction of multi-procedure assembly skills, validated on a real-world robotic assembly platform. Results show accurate sequence learning from a single demonstration, efficient motion planning, and a 93.3% success rate. It contributes to trustworthy and efficient human–machine symbiotic manufacturing systems, aligning with human-centered automation.