Hello everyone

I have a project that deals with multi agent.

I will write briefly about the project and then indicate where I get stuck and what doesn’t work for me

I am currently running the algorithm on one robot

I have a simple robot consisting of a body (center of mass) and 2 legs and its goal is to reach a certain location while navigating a maze that is on a grid.

The catch here is that the legs can only step on the vertices of the grid.

So what have I done so far?!

I used the RRT algorithm and put a bunch of different (random) start and end locations into it

I converted the route obtained from the RRT to a discrete route (which will be the route of the center of mass)

Then I used all sorts of functions and chose for each time step the best point to place each leg on the grid

I created a csv file containing the following:

- Angle from the center of mass to the target (which will actually be a kind of direction vector)
- The angle of the legs (how the legs are positioned in relation to each other)
- 25 binary values that represent the local environment of the robot in the maze (indicate whether there is an obstacle or not)
- The desired angle of the legs position (for the next step)
- dx, dy of the center of mass (that is, how much it must advance in the X axis and Y axis in relation to its current position. It is important to note that this is a relative value and not an absolute position in space

1-3 indicate the input to the network

4-5 indicate the output of the network (the actions)

Because of the problem with using angles (as a result of their periodicity, 0=360 or the distance between angle 1 and angle 359 is not large), I decided to use sin, cos of the angle

That’s why my input is size 29

My output is size 4

It is very, very difficult for me to find a proper prediction of the angles. I get R^2=30% which is very low and does not indicate learning

-I performed normalization, deletion of abnormal values, etc.

(Since there are not many angles of the position of the legs, I tried to divide the problem into 2 - classification for the output of the angles and regression for the output of the positions, but still it does not work well)

Does anyone have an idea what I should do? How to treat data?

Which model to work with? Maybe add other inputs to the network?

How should I treat the angles? I’m sure that’s where the problem is

I get a very bad confusion matrix

Thank you very much everyone!!!