Adventures Building a Self-Driving RC Car | Hacker News Reader

Adventures Building a Self-Driving RC Car

rahulrav.com

173 pointsrahulrav42 comments

If anyone here is interested, check out https://diyrobocars.com or https://diyrobocars.com.au in Australia. I organise the DIY Robocars events in Australia (and we are looking for a sponsor). I'd also recommend joining the Donkeycar and DIY Robocars Slack groups which is where most of the discussion happens in between events.

I got really excited but then the site was a bit of a let down... forms to fill, missing guide... meetup last October...

The Jetson Nano provides a very fair amount of computing power for AI applications like this- I've been following JetsonHacks[0] for a while and there's some neat projects posted on the site. The site owner's Youtube channel[1] also goes into detail on interoperability across the TX series such as cameras, code libraries, and assembling a car based on MIT's open-sourced Racecar[2] built with a TX2.

[0] https://www.jetsonhacks.com/2019/03/25/nvidia-jetson-nano-de...

[1] https://www.youtube.com/channel/UCQs0lwV6E4p7LQaGJ6fgy5Q

[2] https://racecar.mit.edu/platform

JetsonHacks is pretty great.

Your page can't be scrolled with the keyboard and it's missing the native scrollbar. You should remove those overflow rules from the body and .content div.

Apologies. This is a new blog I am working on. Will fix it.

It can, you just have to focus the body element (IE click somewhere on the page) first.

It also breaks the Reader mode in Safari.

If you have even a peripheral interest and are in the Bay Area come to the meetup / race referenced - all are welcome!

It’s surprisingly fun and they serve an awesome BBQ lunch - both me and my 7 year old son enjoyed the program immensely -

https://www.meetup.com/DIYRobocars/

We also have a fun one in San Diego. You can find your local one here (or how to start your own!)

https://diyrobocars.com/local-meetup-groups/

Looks like NYC one might be dead?

https://github.com/markku-ai/donkey-vis

Donkeycar neural network model visualization built to "see" how the AI sees.

Donkeycar is a fantastic project and it got a whole lot better with the recent release of Donkeycar 3.0.

Keep it up guys!

That freaking awesome! I was thinking about making a self driving RC car with image recognition to hunt down some rabbits in my yard, maybe you can help me with that!

Great job man, great job!

I've been thinking about the same thing but for deer. They keep eating my bushes in the middle of the night, so I'd like a little patrol car to chase them away.

I want to make the same thing.

But for humans.

From the Earth.

My mother has found the a motion detector water sprayer works great for this application. Deer comes by and they're sprayed with water.

It looks a lot like this one: https://www.amazon.com/Hoont-Blaster-Repellent-Activated-Spr...

Mine is named Max. He's a good boy. :)

Not to be that guy, but is it still an "RC" car if it's self-driving? :)

It is probably getting commands from a workstation/desktop.

I will take the 'that guy' crown from you. I am thinking what is self driving if the video shows it going off the track and not slowing down for the curves. This reminds me of back in the day with slotted race cars that if you didn't pay attention they flew off the track.

This is not to say that it's not admirable but I was not as impressed as I think I should be when someone is doing self driving.

The rules of the race are that you can't cut corners from inside the track. Going out of the track is a natural penalty, so it's allowed. The car actually slows down a bit.

The ML algorithm being used is called behavioral cloning (and it's cloning me). It's just that I am a pretty rusty RC car driver :)

”Going out of the track is a natural penalty”

No, it isn’t. If the track is concave, driving the convex closure of the track shortens the distance travelled, decreases the amount of turning to do, and no corner will become tighter. That must mean the convex closure is faster to ride.

I guess that’s why the red cones are there. Passing them on the wrong side must lead to a steep time penalty or immediate disqualification.

Also, I would think that, in a really tight turn, you would have to slow down so much to stay on the track that overshooting the turn without braking as much is faster (I think the video shows that in the last corner of the round, at the right edge of the video)

This is ethical application of AI I can get behind

Cool project! Thanks for sharing

Two questions to know a bit more:

- how long did it take you to complete that project?

- what is the rough total cost?

My first car took about a week. The second one took a day. The total BOM was around 400$.

Note: I am not great with hardware.

What took you so long the first time or: did you reuse compiled code the second time?

The first time I was building Donkey for an entirely new (and previously unsupported) SBC - the Asus Tinkerboard S. I was also unfamiliar with how everything was put together.

The second time was a breeze. Jetson Nano runs Ubuntu so getting everything to work was relatively easy.

Good to know, thanks for answering!

Awesome project! Are there build details about getting everything running on the Nano? Do you have a block diagram or interconnect schematic that you could post? Thanks!

Yes. Join the Slack channel. We can help you out if you have questions.

Tiny self driving vehicles could be fun and safe way to transport small things. Think an army of cleaning drones gatherings trash to a collection point.

So a roomba?

More like Wall-E

generic roombas, tangos

Definitely. I just hope licensing and regulation don't make it prohibitive, it would be fun even for personal use as well.

You mentioned

> Having read some amazing books on machine learning ...

Which books did you read and what was your experience with them?

The ones I really enjoyed reading were:

https://www.pyimagesearch.com/deep-learning-computer-vision-...

https://www.machinelearningisfun.com/

and

https://www.amazon.com/Deep-Learning-Python-Francois-Chollet...

my one can be found here its simple and still a work in progress https://github.com/aetar will try to implement yolo3 later project name road trip

Watch Dogs 2

George Hotz, former apple device jail-breaker left Tesla to start working on Self-Driving Cars, the DIY approach. So far he has been training the cars with HD Maps and has had decent progress considering his competition. http://comma.ai

That's a lot of mis-information. Hotz exchanged some tweets with Musk but never worked for Tsla. Comma uses cameras, not HD maps.

Reply on news.ycombinator.com