Machine Learning on Arduino Uno was a Good Idea

The journey of teaching a robot to drive autonomously on a race track!

Tools I use:

Subscriber count at the time of upload: 114 418

[jeffstewart]

This is really exceptional work. I love how your thought process is always generating the next possible improvement, and then you just keep pushing to refine your designs.

[sendhan]

Haven't been doing robotics for a while and this is one of the coolest videos. I got recommended.

[kevinlind]

The most intriguing thing about this according to me is realisation that you could work leanly with data sets. Yes, within a certain dataset eg. maze (the square one for example) you want as many laps as possible, but for industrial purposes keeping the the amount of mazes down when you know what kind of mazes the robot will encounter should also avoid bloating the Arduino with unnecessary data.

Also, truly amazing that you made a robot that could race faster by itself than you could race it manually, just by tweaking the motor speed. That goes to show what machine learning can do in terms of work optimisation, sort of like how the search function on a computer vastly outpaces any human manually searching for a document in an archive room.

Thank you for making a video that clarifies so much with so little!

[patrickjdarrow]

Great project. The reason it was able to handle the increased speed was twofold: 1) the control dynamics were similar enough at both speeds and 2) the sampling rate was high enough that the time delta didn't have an impact on the stateless prediction model.

[matthiasneumeister]

Well done! This was a very satisfying video to watch. Well explained and I totally understand the thrill of building something that actually works in the end!

[TheSelfUnemployed]

On an Arduino! I am grabbing a LIDAR module as soon as possible. Thank you for the videos.

[fireheadpet]

Great video and content! Love how you bring us through the journey of your experiments and the tidbits of discoveries that is available oit there. BTW, little editing features like the ghosting effect really elevates your game. I must agree with other commenter, some of your voice recording suffers in quality (when in testing area, hard walls). It does not impact my opinion but, you are competing for attention against others. I hope you continue to push this project further. Maybe like an iRobot that travels throughout the house for guard duties or identify any new objects...

[midnight]

“I am an algorithm I need more learning and training” ❤ cheers to you!😊

[alexanderyang]

Fascinating! Thanks for posting this video!

[scaletownmodels]

Very cool. I've been programming for a little over 40 years but I've never had the time / opportunity to delve into machine learning. Closest I've gotten is using AI services for photo processing.

Now that it can drive itself it would be an interesting progression to give it memory of where it's been, building a map of the course it travels and being able to use that to plot improved trajectories for future loops.

Just like we're slow when traversing unfamiliar territory but with repeated trips we can anticipate and optimize our course. You should be able to borrow from tech such as CNC path processing which can optimize acceleration / deceleration for curves and apply that to steering. Just an idea.

[JeromeDemers]

Really like the ghost version when comparing the speed. That was nice. Make a other one and see if you can train to pass slow robots.

[simonwatz]

Interesting and very inspiring! Thanks

[gasior]

Great job! It would be great to see this robot learning by itself by reinforcement learning

[edgar]

Great work. You do already amazing things. And you are young. I wonder what you will do in a year or 5 or 10. Keep going!

[HarpaAI]

🎯 Key Takeaways for quick navigation:

00:00 🤖 *Introduction to the robot and project setup.*
- Introduction to a small robot with machine learning algorithms running on Arduino Uno.
- Overview of the project's goal: autonomous navigation on a racetrack.
- Mention of the steps to be covered in the video, including building the robot, creating a racetrack, data collection, processing, and a final race.
02:20 🤖 *Building the robot and the racetrack.*
- Description of the robot's construction using an open robotic platform.
- Explanation of using simple blocks for the robot's chassis and adding necessary components.
- Improving traction on robot wheels with TPU tires.
- Innovative use of cardboard for creating racetrack walls.
05:01 📊 *Data collection setup.*
- Installation of a Bluetooth module and an SD card for data collection.
- Explanation of recording lidar measurements and control labels while driving the robot.
- Details about the data format and collection process.
07:10 🧠 *Processing and training the machine learning model.*
- Discussion of feature selection to reduce data dimensionality.
- Overview of experimenting with different machine learning classifiers.
- Mention of using Python libraries for processing and training.
09:08 🏁 *Testing the robot's autonomous driving capabilities.*
- Introduction to testing the robot's performance on various racetracks, including square and figure-eight.
- Highlighting the ability to adapt to new racetracks with additional training.
- Preparing for a final test on a complex racetrack.
11:40 🏎️ *Achieving high-speed autonomy.*
- Surprising results as the robot successfully handles high-speed autonomous driving.
- Discussion of motor speed settings and PWM signals.
- Comparison of robot performance between manual control and machine learning algorithms.

Made with HARPA AI

[CraigHollabaugh]

This was awesome, congratulations!

I've used o-rings for tires for 3D printed wheels.

[christophersmith]

Watching this, I was shocked to realise that you don’t have a million subscribers. People are missing out!

[OtakuRealist]

Amazing. I like what you're up to. Keep it up!

[majdthabit]

wonderfull thank you for sharing and good luck

[sumitmamoria]

Nice Work! It would be interesting to see if one could "simulate" the movements for a rectangular track, instead of training on the actual path. I would guess, it would lead to comparable results. If yes, then the advantage with simulation is that one can design more complex paths without actually building them - making the training process very efficient and robust.