Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Machine learning of billiard ball dynamics
This modelling project supports the billiard-playing robot at the Automatic Control Lab. It considers the problem of reducing down complex billiard ball physics involving cue ball spin, using machine learning tools and camera feedback.
To start in September 2019 or earlier.
The Automatic Control Lab in D-ITET is building a robot snooker player, with the ultimate aim of challenging the world champion. Currently several students are working on the vision system, artificial intelligence unit, and cue control.
The next step is to understand how to make spin shots, which are important for any strategic player. To do this, it is necessary to characterise the highly complex interactions between spinning billiard balls with a low number of parameters. The simplified model can then be used for control.
The Automatic Control Lab in D-ITET is building a robot snooker player, with the ultimate aim of challenging the world champion. Currently several students are working on the vision system, artificial intelligence unit, and cue control.
The next step is to understand how to make spin shots, which are important for any strategic player. To do this, it is necessary to characterise the highly complex interactions between spinning billiard balls with a low number of parameters. The simplified model can then be used for control.
The project contains several steps:
1) Review a previous project's outputs, which characterised how the 3D physics of "straight" shots deviate from those predicted by a 2D "hockey puck" model of ball collisions.
2) Decide on a set of parameters that characterise the outcome of a spin shot (i.e., one where the cue ball is struck in an arbitrary location and at an arbitrary angle with the cue), both for the cue ball and the target ball.
3) Create controlled simulation tests using the "FastFiz" library, which includes a detailed model of 3D billiard balls' rolling and sliding actions. Use these to fit simpler regression models of the same physics.
4) Use the regression models as an input to shots taken with the robot arm, and validate the parameters from real-world data. Adjust the models so that they can be used to decide on and predict real-world shots.
The project contains several steps:
1) Review a previous project's outputs, which characterised how the 3D physics of "straight" shots deviate from those predicted by a 2D "hockey puck" model of ball collisions.
2) Decide on a set of parameters that characterise the outcome of a spin shot (i.e., one where the cue ball is struck in an arbitrary location and at an arbitrary angle with the cue), both for the cue ball and the target ball.
3) Create controlled simulation tests using the "FastFiz" library, which includes a detailed model of 3D billiard balls' rolling and sliding actions. Use these to fit simpler regression models of the same physics.
4) Use the regression models as an input to shots taken with the robot arm, and validate the parameters from real-world data. Adjust the models so that they can be used to decide on and predict real-world shots.
Joe Warrington and Nikos Kariotoglou, Automatic Control Lab (warrington@control.ee.ethz.ch, karioto@control.ee.ethz.ch)
Joe Warrington and Nikos Kariotoglou, Automatic Control Lab (warrington@control.ee.ethz.ch, karioto@control.ee.ethz.ch)