tlack/PPO-positioner-GM6020
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
Attempt to use reinforcement learning with PPO instead of PID controller to position GM6020 motor via CANBus
2 commits 1 branch 0 tags 35 KiB
  • Python 100%
Find a file
Exact
Thomas Lackner 49aec2ee16
 Update README.md
2020-11-16 21:44:24 -05:00
gm6020.py first 2020-11-16 21:36:26 -05:00
gym_env_gm6020.py first 2020-11-16 21:36:26 -05:00
ppo_positioner.py first 2020-11-16 21:36:26 -05:00
README.md Update README.md 2020-11-16 21:44:24 -05:00
test.py first 2020-11-16 21:36:26 -05:00
train.py first 2020-11-16 21:36:26 -05:00

README.md

RoboMaster-Master

The DJI Robomaster GM6020 brushless motor can report its position faithfully in the range 0..8191 at a rate of 1,000 times a second, but lacks the ability to rotate to a specific position.

Most would use a PID controller to handle that, but I was curious about using reinforcement learning. As of this point, 40k episodes in, not much luck.

This code uses Linux's built in CAN driver. I used a Raspberry Pi 4 and CAN hat.

Contents

  • test.py - manual test script to exercise GM6020

  • train.py - choses hyperparameters randomly from options, dispatches PPO

  • ppo_positioner.py - PPO implementation

  • gym_env_gm6020.py - OpenAI Gym that attempts to locate randomly chosen goal position on motor

  • gm6020.py - DJI Robomaster GM6020 CAN interface library