there is my current version, using MPC and three phase custom FOC control : kzbin.info/www/bejne/npTbdWyPadlresU as first experiment with advanced control, I decide design it as positional control, to avoid kicks there is input shaper before main controller - which is not good solution, better is speed change limited by controller itself ... anyway - there was not tunning required, after dynamics system identification (using EM algo with two kalman filters) the controller can be easily synthetized (by optimization or Riccati eq. solving) I think, PIDs should be avoided in modern world - there is no mathematical reason to use them, they don't match model, they are unable to free manipulate all poles of controlled system, and they works only for single input single output nowadays best way is to use MPC

there is github under video github.com/michalnand/motoko_ice_dragon cost is around 100USD..200USD, however you are never buying one mcu, one resistor, one PCB ... most expensive is time invested - even having that github code, it will never works - you have to tune lot of parameters

@@michalchovanec6022 ohh actually i dont know much about electronics tho, i created some projects but i took help from youtube and chat gpt, but 100, 200 usd is out of my budget

in general way it is filter, "shaping signal" or removing some frequencies from signal - mostly known from 3D printing, where it is basically notch filter, removing dominant resonant frequency, in my case (also generic case of any CNC mechanism), you wanna minimise jerk (3rd derivative of position) - shaper in my case simply limits maximum speed change, it is like ramp, or some S-curve : google for jerk, or S curve motion profile - implementations can vary, but I'am sure even simple linear filter can do a lot of improvement github.com/michalnand/motoko_ice_dragon/blob/main/firmware/LibsCommon/shaper.cpp

@@michalchovanec6022 thanks a lot, it's amazing!

inov 8 roclite is my choice, www.inov-8.com/running/shoes/mens for Teide hikes I suggest longer merino socks, and warm clothing

Iam using mcu stm32f765 (cortex m7, 216MHz), enough power for real time MPC control, recurrent neural network planning, and FOC for two motors there is programming connector, both, st-link and uart bootloader works

@@michalchovanec6022 cool, thanks for letting me know

this robot have this slow motors : www.ebay.com/itm/123945150416 good for run in maze, or line follower competition where are sharp curves (radius less than 5cm) for high speed tracks, you can use any out runner motor only important is to figure out how to mount magnet on shaft, for encoder

it is bit old project ... I finished exporting trained models from pytorch into int8 github.com/michalnand/classification/tree/master/embedded_inference also watch my robot, using CNNs on 72MHz ARM cpu : kzbin.info/www/bejne/roake2SCdpuLrcU

I think this is usable only for few parameters system - maybe for some basic PID tuning ... for neural network we have ADAM solver

Iam old man with pure memory :D

this is old project which is obsolete now ... my current project is here : kzbin.info/www/bejne/e2qpe3x6pNyBo7c hackaday.io/project/163799-motoko-uprising-deep-neural-network-line-following sources : github.com/michalnand/motoko_uprising

thanks Iam working on mapping the line - neural network will be used for line type classification and map creating; on second run I can use particle filter (or some other similar bayess filter) and map for much faster run

line following is simple task - robot is in real time, arround 20Hz samplig (robot is slow), sparse distributed memory allows one shot training

horolezeckametodika.cz/ucebnice/slanovani/slanovani-s-osmou

Michal Chovanec aha dik

I want to make one please help..

this is old project - now I have better device, but I don't have time to make some video, or post ... you can start with some basic controll theory : PID, filter, Z-transform - but no in matlab, try to program it from scratch, in C++ or python ... then you can learn some basics using machine learning - there are lot of frameworks (caffe, tensor flow ...) but to know how to tune it, you need to know how neural networks works, and this is for 3..5 years of hard work study; you can find some inspiration in my recent post, how to fit neural network into arduino : hackaday.io/project/41159-deep-neural-network-on-arduino-mnist-handwritten

Which microcontroller uses

there are 8 rgb sensor apds9950 or apds9960, because they have the some i2c address Iam using bit banding i2c : 8 sda channels and common scl

it is rare map from local retired miner - I own only copy, but not to share with unknown guy