motoko uprising new

motoko uprising line following robot,

highlights

• cpu stm32f7, 216MHz ARM Cortex M7

• DRV8834, TI motor driver

• 8x phototransistor + white led as line sensor

• IR leds for obstacle detection

• 2x pololu motors HP, 1:30 with encoder

• expansion : camera, laser sensor

• custom operating system

• reccurent LSTM in controll loop

• total 3 other PIDs used

structure

• src

  • firmware - robot main firmware
  • robot_debug_py - debug application
  • AI gym env for reinforcement learning gym-line_follower
  • TODO rl_line_follower rl_line_follower

• hardware

  • eagle files hardware
  • gerber outputs files
  • stl for 3D printed parts files

supervised learning for line position prediction

key points

• input is time series sequence from 8 line sensors

• output is classification into three classes (left, straight, righ)

• RNN GRU with 64 units used, and linear layer with 3 neurons

• confusion matrix result for float32

accuracy   = 89.037 [%]
hit_count  = 116703
miss_count = 14369

class_accuracy = 90.952%   90.474%   84.811%   


confusion_matrix = 
       28326        1376           8
        2817       57907        5448
           0        4720       30470

• confusion matrix result for int8 quantization

testing result
accuracy   = 87.792 [%]
hit_count  = 115071
miss_count = 16001

class_accuracy = 79.176%   91.345%   87.918%   


confusion_matrix = 
       22858        3566           0
        6011       62800        4041
           0        2383       29413

reinforcement learning in Pybullet ENV

key points

• state is 256 last line positions, motors velocities, and line present flag

• PPO algorithm for continuous action space

• GRU model with three output heads (mean, variance, value)

hardware