Racing the physical AWS DeepRacer is not easy to organise in 2020. Thankfully Jon Myer and David Smith have stepped in to provide the entertainment and a learning opportunity. As a re:Invent special you can win a DeepRacer Evo! Read on for the details where and when the next race will take place.
You don’t need the physical DeepRacer to start training and racing in competitions, just head to the AWS DeepRacer Console and get started. But there is so much fun in hitting a real track and competing against others!
As we wait for it to be safe to return to AWS Summit races and community meetups, Jon and David with their helpers have worked hard to bring the experience to you. Every other week Jon and David Stream a live event in which racers can try their models on a physical car and on a physical track.
It works this way: you submit a model through a community race to make it into the live stream, then Jon, David or any of their pit crew members will load the model onto a car, put it on a track, chase it around and put back on track if needed, control the maximum speed on it to try and help it do its best, and time your laps.
How can I compete with the pros? They are fast
Important element of DeepRacer is learning, that’s why DBS had over 3000 of their employees compete internally, and similarly Accenture, DNP and many others.
To make it simpler for newer racers to make it to the race, Jon and David set up three categories:
- Rookie race – a standard time trial category for the new-starters and less experienced ones
- Pro race – another time trial, but here you’ll need to up your game a bit
- H2H race – the ultimate challenge where you don’t only have to stay on track and go fast, but also avoid the opponent (or maybe get them to crash out of the race?)
This is to ensure that everyone gets their time. Races are taking place regularly and so you will get to iterate on your learnings and try to improve. You can also just keep submitting your existing model.
How do I take part?
To see your model raced on a track you need to:
- Train your model here (see this article on how to get started by AWS ML Hero Juv Chan)
- Submit it to a community race announced in the AWS ML Community #dr-underground channel: http://join.deepracing.io
- Get a lap good enough to be in the top group
- Send a downloaded model to either Jon or David in Slack when they ask you to
- Join the live stream on Twitch and bring some pop corn
That’s it! Dates and links for the next race are available below.
AWS re:Invent Special
There are many DeepRacery goodies in store for you at re:Invent 2020 (read here), and here’s one more:
During DeepRacer Undergroud on the 9th of December you can win a DeepRacer Evo in each category!
And a re:Invent Special Special: If we manage to beat 50 entries in any of the categories, AWS will throw in an extra Evo for that group!
Just remember that in December you can train DeepRacer in the console for only $1 per hour of training.
DeepRacer Underground re:Invent Edition
The next race will take place on the 9th of December at 9am PST (5 pm UTC) and can be watched on Twitch.
Track: bot community races and the underground one will take place on re:Invent 2019 track.
Time to submit a model to the community race ends on the 7th of December at 9am PST (5 pm UTC) – top 3 qualify:
Time Trial Rookie Race: join here
Time Trial Pro Race: join here
EVO H2H Racer: join here
Tips & Tricks
A few things to take into consideration when planning for the DeepRacer Underground:
Physical racing is different to virtual racing. simulation is a perfect environment. You don’t get shadows, the car doesn’t shake, the input image is always sharp etc. It’s not exactly that but if you get your phone and start recording and run for a bit, you will see a stable, clear image and will recognise everything on your way but the recording will be shaky and blurry. That’s the kind of difference between what simulation provides and what the car sees. This has quite significant consequences: when training for the simulation you will tend to overfit and crank the speed up as much as possible, in physical it’s better to train for stability and generalisation. Sumil offers a few insights into this aspect:
Also the physical car doesn’t take speed values into account as much as speed levels (so fastest speed in the action space is speed 100%, and the second one is 50% regardless of whether it is 2 m/s and 1 m/s or 1 m/s and 0.5 m/s).
Don’t overtrain. for physical track I’ve found that cars do better with models trained for a shorter time.
You can submit a different model in community race and on track so you can apply different strategies for community race and for the physical one 🙂
You can get lots of help in the community: http://join.deepracing.io