A Task-Learning Strategy for Robotic Assembly Tasks from Human Demonstrations.

Area of Science:

• Robotics
• Artificial Intelligence
• Manufacturing Automation

Background:

• Traditional robot programming is inefficient for human skill transfer.
• Current methods lack flexibility in learning and generalization.
• Need for intuitive human-robot collaboration in manufacturing.

Purpose of the Study:

• To develop a novel task-learning strategy for robots.
• Enable flexible skill acquisition from human demonstrations.
• Facilitate skill generalization to new task scenarios.

Main Methods:

• Markerless vision capture for human hand movement acquisition.
• Heuristic segmentation to define movement primitives (MPs).
• Gaussian Mixture Models (GMM-GMR) and Dynamical Movement Primitives (DMPs) for trajectory learning and generalization.
• Visuo-spatial skill learning (VSL) for goal configuration learning.

Main Results:

• Successful segmentation of human movements into MPs.
• Extraction of optimal trajectories using GMM-GMR.
• Effective trajectory generalization via DMPs.
• Accurate learning and generalization of goal configurations using VSL.
• Demonstrated exact pick-and-place points and smooth, human-like trajectories in peg-in-hole experiments.

Conclusions:

• The proposed task-learning strategy enhances human-robot skill transfer efficiency.
• Robots can flexibly learn and generalize skills from limited demonstrations.
• The method enables robots to adapt to new task situations effectively.