Learning Trajectories for Robot Programing by Demonstration Using a Coordinated Mixture of Factor Analyzers.

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Humanoid robots require sophisticated control for complex tasks.
• Learning from demonstration (LfD) is a key paradigm for robot skill acquisition.
• Existing LfD methods often struggle with high-dimensional and redundant robot systems.

Purpose of the Study:

• To develop a robust method for learning humanoid robot trajectories from human demonstrations.
• To represent robot motion as a sequence of motion primitives within a nonlinear dynamical system framework.
• To enable robots to synthesize and execute learned trajectories accurately.

Main Methods:

• Utilized a hidden Markov model (HMM) to learn the probability of residing in each motion primitive.
• Employed a coordinated mixture of factor analyzers as the emission probability density for the HMM.
• Developed a dynamic system acting on a shared manifold for motion synthesis between human and robot.
• Performed stability analysis to ensure robustness to trajectory deviations and transitions.

Main Results:

• Successfully synthesized motion for a 19 degree-of-freedom humanoid robot.
• Demonstrated robustness to deviations and transitional motion between learned manifolds.
• Reduced model complexity for kinematically redundant robots.
• Showcased reduced observation requirements for effective learning.

Conclusions:

• The proposed HMM-based approach effectively learns robust humanoid robot trajectories from demonstration.
• The method offers significant advantages in model complexity and learning efficiency for redundant robots.
• Experimental validation confirms the system's capability to synthesize and execute human-like motion.