General robot kinematics decomposition without intermediate markers.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Serial manipulators with many degrees of freedom present calibration challenges.
• Existing machine learning calibration methods can be complex and time-consuming.
• Current decomposition strategies have limitations regarding pose capture and robot constraints.

Purpose of the Study:

• To present an alternative decomposition strategy for calibrating complex serial manipulators.
• To reduce the complexity and time required for robot calibration using machine learning.
• To overcome limitations of existing decomposition methods.

Main Methods:

• Developed an offline training algorithm for sequential learning of composite subchains.
• Introduced a second training scheme for simultaneous and online learning of composite chains.
• Utilized parameterized self-organizing maps and Gaussian mixture models (GMMs) for simulations.

Main Results:

• The proposed decomposition strategy effectively calibrates serial manipulators without requiring simultaneous end-effector pose capture.
• A twofold decomposition reduced the number of samples needed for calibration to twice the square root of the original number.
• Both offline and online training schemes demonstrated the approach's correctness.

Conclusions:

• The novel decomposition method significantly enhances the efficiency of machine learning-based robot calibration.
• This approach offers a more flexible and less constrained alternative for calibrating high-degree-of-freedom serial manipulators.
• The findings pave the way for faster and more accessible robot calibration.