Skip to content

INTELYMEC/deepPID

 
 

Repository files navigation

Deep PID

This is the implementation of the inverted deep PID algorithm found in the article:

"An adaptive deep reinforcement learning approach for MIMO PID control of mobile robots"

submitted to ISA Transactions, currently under review.

This repo contains:

  • deep PID controller using tensorflow
  • Vanilla DDPG
  • DDPG with inverted gradients
  • An implementation of the TD3 algorithm

Requirements:

How to run

Have in mind that the Pioneer simulator speeds up the simulation time, so you will need a machine that is able to run it. To run the Deep PID with the pioneer with the inverted gradient:

python main.py --alg iddpg --gpu gpu --epochs 1000

The alg argument selects with which algoritm to train the agent, iddpg or it3.
Results on the real robot can be obtained such as:

alt text

The algorithm periodically saves the actions and velocities obtained every x number of epochs, this number can be changed by using the argument save_mod. The neural networks are saved automatically after the max number of epochs is achieved.

DDPG demos

All the implemented algoritms can be runned as demos with different gym environments. All the implementations are within the classes files. So for instance running:

python ddpg.py 

will run the ddpg algorithm using the inverted gradients for the gym's pendulum example. While td3.py runs the same for the td3 algoritm.

Authors

Ignacio Carlucho, Mariano De Paula, Gerardo G. Acosta

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%