Design a reinforcement learning (RL) based controller to stabilize a quad copter

Hi everyone, hope all is well at your end.

I need to design a reinforcement learning (RL) based controller as an asignment to stabilize a quad copter from an initial position, orientation and angular velocity. The way I look at the problem is that due to continuous action space i intend to apply a policy gradient actor critic algorithm of RL. Request guide on some Matlab code that utilises policy gradient actor critic algorithm of RL that would help me solve this problem.

Although i am developing a simple quad copter model but if anyone can help me with an existing model that will validate my case ill be grateful for that as well.

Thanking in anticipation...

Regards

[Naveed Nekoo]

Any update on this?

Answer 1

[Emmanouil Tzorakoleftherakis]

Hello,

Not sure if you are still looking for a solution, but starting in R2019a, you can do deep reinforcement learning directly in MATLAB and Simulink with Reinforcement Learning Toolbox.

This video shows how to use the toolbox to teach a robot how to walk with continuous actions. You could use this as a starting point and make changes as necessary for your quadcopter application.

You can also use some existing quadcopter models if you don't want to build your own from scratch. See for example the following links:

https://www.mathworks.com/matlabcentral/fileexchange/63580-quadcopter-simulation-and-control-made-easy

https://www.mathworks.com/examples/aeroblks/mw/aeroblks_product-asbQuadcopterExample-quadcopter-project

https://www.mathworks.com/videos/drone-simulation-and-control-part-4-how-to-build-a-model-for-simulation-1539585112546.html

Hope this helps.

[Ali Aboubih]

I have been attempting to get this to work on my own for a while, and have tried various agents and network architectures, but I haven't had any luck. Would you be able to assit me? Here is my code for Actor-Critic, the simulink file, and I still get results like this:

[Unmanned Aerial and Space Systems]

Hello, did you solve your problem?

[Muhammad]

Dear Ali, i believe that reward function is also very important for proper training and convergence to reference values. The proper selection of reward function will help you to get good results.

You need to think of the reward as a useful optimization criterion for your problem.

I hope with proper reward function your problem be solved.