Geometry-Constrained Learning-Based Visual Servoing with Projective Homography-Derived Error Vector

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Camera-in-hand visual servoing traditionally requires camera intrinsic parameters, depth information, and robot kinematic models.
• Existing model-free methods often struggle with computational complexity and robustness to feature occlusion.

Purpose of the Study:

• To develop a geometry-constrained, learning-based visual servoing method that eliminates the need for explicit camera and robot models.
• To enhance the robustness and learning efficiency of image-based visual servoing systems.

Main Methods:

• Utilizes a cerebellar model articulation controller (CMAC) for online Jacobian estimation.
• Introduces a fixed-dimension, uniform-magnitude error function based on the projective homography matrix.
• Incorporates geometric constraints (e.g., collinearity preservation) into the neural network update process.

Main Results:

• Achieves robustness against feature occlusion by not relying on individual feature points.
• Reduces computational complexity through a constant Jacobian size.
• Demonstrates superior robustness and faster learning rates compared to existing model-free visual servoing methods in experiments and simulations.

Conclusions:

• The proposed method offers a simplified and more efficient approach to camera-in-hand visual servoing.
• Geometry constraints ensure physically plausible control outputs, improving reliability.
• This novel technique advances model-free visual servoing capabilities in robotics.