Learning the Inverse Dynamics of Robotic Manipulators in Structured Reproducing Kernel Hilbert Space

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Accurate modeling of robot inverse dynamics is crucial for control and motion planning.
• Existing methods like rigid-body models are restrictive, while learning-based models demand extensive data.
• A gap exists for efficient and generalizable inverse dynamics modeling without prior kinematic information.

Purpose of the Study:

• To propose a novel approach for modeling inverse dynamics of holonomic rigid-body robots without prior kinematic information.
• To develop structured kernels capable of autonomously converging to rigid-body models while minimizing sample requirements.
• To enhance modeling efficiency and accuracy for general rigid-body robots by incorporating a semi-parametric framework for friction.

Main Methods:

• Embedding robot dynamics within a reproducing kernel Hilbert space using structured kernels.
• Developing a semi-parametric framework that combines structured kernels with parametric basis functions for friction.
• Validating the proposed scheme through simulations and experimental robotic system tests.

Main Results:

• The structured kernels demonstrated autonomous convergence towards rigid-body models.
• The proposed method required significantly fewer training samples compared to traditional learning-based approaches.
• The semi-parametric framework effectively modeled general rigid-body robots, showing improved accuracy.
• Models incorporating the structure of function space yielded higher precision in simulations and experiments.

Conclusions:

• Structured kernels offer an efficient and data-frugal solution for inverse dynamics modeling of rigid-body robots.
• The integration of structured kernels with parametric friction models enhances the generalizability and accuracy of robot dynamics modeling.
• This approach advances the field of robot control by providing a more robust and adaptable method for understanding robot dynamics.