Home > AWS, re:Invent, Serverless > AWS re:Invent 2018: DeepRacer Workshop – AIM206

AWS re:Invent 2018: DeepRacer Workshop – AIM206

November 29th, 2018

20181129_025938202_iOSDe Clercq Wentzel from AWS

Well, we were super lucky. I was excited before as you can’t get more hackathon than a Robocar Rally which this was in the session list as. During the Andy Jassy keynote this morning, DeepRacer was announced which is a machine learning car.

I’m the first to admit I haven’t done much machine learning so this appealed to me as it was for developers with no prior ML or robotic experience.

You can play along: https://github.com/aws-samples/aws-deepracer-workshops

We joined up pit crews, I teamed up with fellow UK VMUGer and Virtual Design Master winner Chris Porter (I’m not ashamed to grab onto the knowledge coattails when needed!).

The idea was to use machine learning to train the car in a virtual racetrack built in RoboMaker to learn to stay on track. One the training model was done, the data could be downloaded to the actual car and then the car would attempt to race a real track using the learning model from the training data.

Image result for deepracerStraightTrack

Reinforcement Learning

De Clercq (originally from South Africa! Smile) went through an overview of reinforcement learning which is the subset of machine learning which the car uses. , You try to train a model by looking at the consequences of the actions taken. A human would bump their head against something and learn not to do it again. Reinforcement learning is getting a machine to do the same thing

The robot car will go around the track and try to maximise the rewards it gets, taking images on the road and working out the best next action to take.

20181129_030255227_iOS 20181129_030419709_iOS

20181129_030458942_iOS 20181129_030618696_iOS

We needed to decide as a group how we would reward our car for getting things right. Staying as close to the central line was a good reward but you could get more sophisticated with working out which waypoints you get to, which direction are facing and whether there is a straight ahead for example. You could decide how close you actually want to be to the center and what leeway you want to give it. Do you penalise the car for going too slow or steering too erratically?

You then start your model and can actually watch the car “driving” in the simulator.

good_sim_training2.png

The model training can take some time, it was suggested 150 minutes but that’s then for each interaction of your model, you can then tweak the rewards you want to give it.

You then need to take your model and put it in the real world. You obviously want your training track to be the same as your real track or your training data will not be effective.

We then headed to the MGM Garden Arena where they had set up a whole race environment. 4 tracks where you could load your models to a car and see how you went. All I’ll say is we obviously didn’t have enough time to train our models!

20181129_050444386_iOS 20181129_052342634_iOS

20181129_050918334_iOS 20181129_050945579_iOS

It was a fun and immersive experience. We had to build, train, and race a robo-racing car.

Oh, and we all received one as well! Smile

Comments are closed.