juangallostra / messenger-basketball-reinforcement-learning Goto Github PK

Complete the info in the README about the different modules implemented for the project.

Short description of the issue

Instead of sampling the basket and ball position every frame they are inside their respective regions of interest some sort of sampling period for the positions should be included as a function of the grid size.

Relevant notes

This should be implemented because if not, although the basket might be moving, if it remains for a certain amount of time inside the same cell of the grid then the series of states used as input for the learning algorithm might yield the wrong conclusion that the basket is not moving.

Organize the repo, update the readme and release its first version.

This issue also implies addressing:

• #18
• #19

When computing the joint angles via the inverse kinematics model the quadrant the angle lies into should be taken into account and checked.

Tesseract is the OCR engine used to detect the current score and it takes quite a long time to process a frame. This reduces significantly the performance. So the score should be computed only when it is strictly required.

So the question is, when is it strictly necessary to compute the score?
Only after a throw, and to deduce that a throw has taken place:

• The ball and basket are being detected after a period of not being detected.

To be able to properly drive the robot to the desired position the servos should be first calibrated.

It would be nice to add a functionality that shows the Q matrix in real time to see how it updates.

Short description of the issue

From time to time the score is read wrong. This might be just something of the video currently being used but it would be worth to devise a way of making sure that the returned score is actually correct. It might help that the expected score is the previous score plus 1, and a check along those lines could be implemented.

Relevant notes

Since in the end we only need to identify correctly 0 (so as to give a negative reward), this may not be an issue at all by assuming anything different than 0 as a successful throw.

In order to obtain the inverse kinematics of the robot the required parameters have to be measured. Furthermore, the measurements should be as accurate as possible.

Add the video or a gif of the fnal result where the robot playing is seen.

This issue refers to solving:

• #9
• #8
• #11