Of interest , the DeepRacer does reinforcement learning while donkeycar has a completely different approach (supervised learning). It would be interesting to see them compete
This is incredibly cool, and I'd love to enter, but despite having one of the races in Sydney, the actual DeepRacer hardware doesn't ship outside of the United States. Baffling from a company that has a local presence in Australia.
If anyone from AWS is reading, I'd love to buy one!
I built a tensorflow model from that trained on a 750 TI for the Udacity Self-Driving Car Engineer Nanodegree (it was part of the course to build such a thing that ran on a virtual track), and after training it did quite well - at least on a virtual track.
A few tweaks to the model, maybe using the TensorFlowLite package or something on the phone - probably work fine for 99.9% of the runs...
I actually completed the Udacity SDC course as well and implemented the same Nvidia NN.
I looked into building my own implementation, but decided against it. I'm not particularly interested in the hardware, just the algorithms that drive it, and the amount of time spent getting the hardware working (not to mention a simulator ) would have been nontrivial.
I'm not with Amazon and as much as I'd like to do the same, I suspect not.
The whole idea of the competition would be to use the same hardware so it's a level playing field from a hardware perspective which turns it into a software competition, right?
Is it hotlapping or a race? One is a lot easier than the other as evidenced by codemasters's valiant attempts to write AI over the last few years.
Also, it looks like they don't know the track beforehand but they seem to be taking absolutely bizarre lines through bends (Or is that an artifact of the model used?)
There is already FIRST robotics competitions for that. FIRST tech league in particular is pretty software oriented. There programs are a great way to get students into programming and learning about writing real software.
There are a bunch of events before they actually ship. They gave a bunch of them out at this past year's re:Invent, so I guess only those people who got them can participate until then.
Udacity SDC alum here as well and this looks awesome. I'm tempted to buy one after reading about DIYRoboCars. I've been thinking about porting openpilot to my Jetson TX2 but have been reluctant since I have no way to test it. DeepRacer seems much funner and safer. I wonder how hackable the car will be. How hard would it be to set up a remote ROS nodes rather using AWS services?
The models should be limited to within the confines of the s3 location you specify (and the physical deepracer device) unless you explicitly permit otherwise. Nothing that I see in the terms of use [0] cover access to the model.
Now that nvidia has released a cheaper jetson board, I'm really itching to get it and build a small RL car. Training on the device would be out of the question though.
It's not clear to me why deep learning is required here. It seems like some dead simple lane following code would probably work just fine. This is generally how these competitions are done.
I got a deep racer at Re:invent but I'm still waiting to get into the preview to play with it. Anyone from AWS cruising the comments here who might be able to enable it for my account?
darn. the window for $249 deepracer is over? Still mentioned as $249 'introductory discount' at https://aws.amazon.com/deepracer/pricing/ but upon clicking through to Amazon.com it's $399 now.
My reading is that pre-order is more expensive than launch price. "Can" vs "will be"?
"You can pre-order your AWS DeepRacer car on Amazon.com for $399. For a limited time at launch AWS DeepRacer will be available at an introductory discount price of $249."
You require AWS to at least activate and set up DeepRacer [0].
You don't strictly speaking need the cloud for training a DeepRacer model, although cloud access (via the DeepRacer console) could be the path of least friction.
Much of the source code for training DeepRacer on AWS can be found on the amazon-sagemaker-examples github repo [1].
The sample training environment relies on Gazebo [2] which, while freely available, has a managed AWS offering [3]. But ultimately, you should just be able to deploy an RL Coach model to the device [4].
For some reason most robotics contests seem to be restricted to highschool students. It's bizarre. It's like "Hey kids get into STEM because robots are cool! OK now stop and go build accounting software."
Edit: Thinking more about it, it's even worse. "Hey kids, robotics is a toy for kids to play with."
I think a big part of it is because as an adult with the time, inclination, some money, and programming skills you're expected to just go off and build things.
I guess the same holds true for other adult things like paint-and-sips (you drink and paint a picture) though, so maybe you've come across an untapped market for bot-and-booze.
Just wanted to compliment you on your self-driving car design. Also skimmed your blog and paper. My intended area of focus is generative modeling but I thought your project was awesome nonetheless.
I'm more of a GCP fan since they provide so much in GCP credits. Unless you're doing hundreds of hours of compute, I don't think that getting a relatively high AWS bill to your card will be a problem.
Access to the DeepRacer section of the AWS console hasn't been made public, and I don't think there's any other way to program it.