Segment, Compare, and Learn: Creating Movement Libraries of Complex Task for Learning from Demonstration

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Learning from Demonstration (LfD) relies heavily on motion primitives, which are often manually generated per task.
• The manual generation of motion primitives is time-consuming and limits the scalability of LfD.

Purpose of the Study:

• To develop an automatic and unsupervised algorithm for segmenting robotic tasks and generating libraries of reusable motion primitives.
• To reduce the manual effort required for acquiring robotic skills through Learning from Demonstration.

Main Methods:

• Initial task segmentation using a heuristic method.
• Probabilistic clustering of segments with Gaussian Mixture Models.
• Grouping similar segments using Gaussian Optimal Transport on Gaussian Processes (GPs) to compare movement structures via energy cost.

Main Results:

• Successfully generated libraries of motion primitives by autonomously segmenting tasks into common subtasks.
• Validated the algorithm's effectiveness in real-world robotic manipulation tasks.
• Demonstrated superior or comparable performance against state-of-the-art algorithms.

Conclusions:

• The proposed algorithm enables efficient and autonomous acquisition of robotic skills through unsupervised task segmentation and primitive generation.
• The method supports multimodal information and requires no prior task knowledge, enhancing its applicability.
• Generated motion primitives effectively encapsulate movement structures, facilitating subsequent learning processes.