Phase portraits as movement primitives for fast humanoid robot control.

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Traditional robot control relies on computationally intensive optimal control, sensitive to model accuracy.
• Animals exhibit efficient, adaptable motor control with natural inference of dynamics and coordination.

Purpose of the Study:

• To develop fast, computationally light robot controllers using a motor skill learning perspective.
• To introduce Phase Portrait Movement Primitives (PPMP) for autonomous learning under mild modeling assumptions.

Main Methods:

• PPMP predicts dynamics in a low-dimensional phase space using coupled oscillators for phase prediction, replacing model-based state estimators.
• The control policy is trained by optimizing oscillator parameters linked to a kinematic distribution (phase portrait).
• Demonstrated on a 20-DOF humanoid upper body for tasks requiring fast reactions and anticipative pose adaptation.

Main Results:

• Achieved efficient training and execution of PPMPs on a complex humanoid robot.
• Enabled fast reaction times and anticipative pose adaptation in both discrete and cyclic tasks.
• Showcased reduced computational load and improved performance compared to traditional methods.

Conclusions:

• PPMP offers a computationally efficient and effective alternative for fast robot control.
• The motor skill learning approach, via PPMP, enables autonomous learning and adaptability in robots.
• This primitive successfully addresses challenges in high-dimensional robot control and dynamic interactions.