Autonomous Golf Putting with Data-Driven and Physics-Based Methods

TL;DR

This paper presents a self-learning golf robot that combines data-driven and physics-based methods, using a neural network and physical modeling to autonomously learn putting from arbitrary positions on the green.

Contribution

It introduces a novel integration of data-driven neural networks with physics-based modeling for autonomous golf putting, reducing real-world interactions needed for learning.

Findings

Successful demonstration of autonomous putting from arbitrary positions

Effective combination of neural network and physics-based modeling

Reduced need for extensive real-world training interactions

Abstract

We are developing a self-learning mechatronic golf robot using combined data-driven and physics-based methods, to have the robot autonomously learn to putt the ball from an arbitrary point on the green. Apart from the mechatronic control design of the robot, this task is accomplished by a camera system with image recognition and a neural network for predicting the stroke velocity vector required for a successful hole-in-one. To minimize the number of time-consuming interactions with the real system, the neural network is pretrained by evaluating basic physical laws on a model, which approximates the golf ball dynamics on the green surface in a data-driven manner. Thus, we demonstrate the synergetic combination of data-driven and physics-based methods on the golf robot as a mechatronic example system.