Robot Task-Constrained Optimization and Adaptation with Probabilistic Movement Primitives.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Robots in manufacturing require skill learning and adaptation for efficiency.
• Movement Primitives (MPs) are key for encoding robot skills.
• Probabilistic Movement Primitives (ProMPs) offer enhanced flexibility through movement distributions.

Purpose of the Study:

• To develop a method for learning MPs from limited human demonstrations.
• To enable robots to adapt learned skills to diverse task constraints like waypoints, joint limits, and obstacles.
• To improve robot adaptability and efficiency in manufacturing.

Main Methods:

• Learning Probabilistic Movement Primitives (ProMPs) from a small number of human demonstrations.
• Utilizing an improved via-point generalization for smooth trajectory generation.
• Employing a task-constrained optimization method incorporating constraints analytically into a probabilistic framework.
• Minimizing Kullback-Leibler (KL) divergence using gradient ascent-descent for ProMP optimization.
• Developing a unified robot movement adaptation method for single and multiple obstacles.

Main Results:

• Successful learning of robot skills from minimal human input.
• Effective adaptation of movements to various task constraints including dynamic obstacles.
• Demonstrated generalization of smooth trajectories with encoded ProMPs.
• Validation of the approach on a 7-DOF Xarm robot through extensive experiments.

Conclusions:

• The proposed method enables robots to learn and adapt skills efficiently from human demonstrations.
• This approach enhances robot flexibility and applicability in complex manufacturing environments.
• The framework provides a robust solution for task-constrained robot movement adaptation.