AWS DeepRacer Vegas Finals
Placing 4th In The World Championships
Summary of this article;
- Lots of pics and a few youtube links
- Before we start
- DeepRacer
- Preparing for Vegas
- On The Road
- Race Day-An epic tale of ‘almost glory’, the highs and lows
- Prepare for the press
- Impressions of the competition
- Win or lose, enjoy yourself and learn how to improve
- Plans for 2020
If you haven’t read Glenn Horan’s origin movie version covering the run up to re:Invent I would strongly encourage you to do so.
Before we start
There has been massive support from Liberty Mutual/Liberty IT and my Quantum unit from the start of our journey to the re:Invent Championships, from running the whole competition to sponsoring our wildcard race entry/conference attendance passes after we had won.
Colleagues from home and the US also came and supported us on our race days in person. Having the crowd behind me before and during the race was very encouraging, thank you everyone.
DeepRacer
If you’re not familiar with AWS DeepRacer, it’s the “World’s First Autonomous Racing League”. Small scaled race cars programmed by python and taught using Reinforcement learning, a branch of AI Machine Learning.
You can read through my previous 5min Medium article explaining more about DeepRacer and the Liberty League.
Or if you prefer the excitement of a main staged summit race then watch some DeepRacer TV, (I recommend the Tokyo episode).
Preparing for Vegas
Build up the hype in the office, talk the talk. We hoped our knowledge would inspire others and that we could live up to Outrun’s expectations.
We also offered attendees the chance to make any models they wanted and we would test them in Vegas, if their model won, 100% credit would go to them.
Getting a DeepRacer car delivered to the office
(A prize from winning the Liberty League)
Setup said car at home to test if it can connect to network and be calibrated correctly on the console web app.
Next I got my DeepRacer models ready, checked the logs and made the best decisions I could for getting complete laps as fast as possible. The below is not a good model example but that’s 4th place for ya ;)
Saved £1486 by spending £14 on white tape and generally guess at the shape of the track we planned to race on. Kudos to David Fyffe for spending a lunchtime and more on taping it together with me.
I spent many early mornings, lunchtimes and after work sessions testing out models and tweaking with the car’s steering and speed calibrations. The car had immense trouble at the start even getting around the track, there was a lot of distractions that didn’t exist in the simulator. But I persevered and matured the models I had across different simulator tracks and reward function tweaks that focused on stability and reduced steering. Eventually I was able to get a handful of models able to complete 1–2 laps, they weren’t impressive times, 16–22 seconds, but time was up and they would have to make do. I was just hoping the real track and its environment would allow me to speed up to get faster laps.
So preparation was done, with roughly 30 models armed to the USB, we where ready to test on the real track.
That would have to wait until we got to MGM Grand Arena in Vegas.
On the road
It took a lot of travelling, (over 24 hours from door to hotel lobby), so it was important to keep morale high and keep our minds off the race.
Although some folks couldn’t handle the pressure and tried to escape.
But it all worked out in the end and the journey was underway!
We put some early practice into our podium poses.
Have to act the part you want, right?
The eve of Race Day saw team Outrun getting all our models collated together and getting ready for an early start to test as many models as possible before our official race time around 3pm.
RACE DAY (An epic tale of ‘almost glory’, the highs and lows)
After 1000% confirming I was INDEED racing it was time to get my head in the game, with the team behind me I knew we could go far, I just needed to actually SEE a model run on the real track.
Right after the below pictures we tried to get another racer into the Championship as there was 3 qualifying spots left. However it wasn’t in the cards even though the guys raced the best they could. A good lesson in DeepRacer and possibly ML in general; nothing is a guarantee in terms of repeatability.
Expect to have events happen that you need to adapt too, i.e have backup models, switch to them early, try much higher and lower speeds to explore the models you are currently racing.
4 minutes pass very quickly but you can achieve a lot if you have a timeframe in mind and a plan if things go south.
So I got down mega early to the Arena the next day, (6.45am for an 8am opening) to give my models a test in “Open Play” available to the public.
David Fyffe and Glenn Horan joined me to test our models on the real tracks and early results were promising. With us setting under 14 seconds, then 12, then finally under 10 seconds sitting at second place for the practice sessions.
Now we started to look competitive, were all the previous efforts starting to pay off? Obviously the title of this article answers this question, but let’s carry on like you didn’t read that.
Moving onto my group of 16, only the top 4 were allowed through. After some good races and some bad we managed to finish top of our group, making the next day a bit easier to start off in as day 2 was seeded.
And this was the final standings across all groups at the end of day 1.
On Day 2-The high and the low
Now we needed to prove our stuff, but only 3 times. 3 wins in a row would get us to the final. In our first race we managed to break the world record with an 8.9 second lap and claim the first of 3 needed wins.
Video of breaking the world record (Not sure why it doesn’t embed)
https://www.youtube.com/watch?v=IRwklHwm9yg&list=PL0vhBOrwsa9E09tZfIebGzNimSI_5AH0E
This was our highest, who could stop us?! Well 50 minutes later our world record was broken again with a 8.5.
Then again with a 8.4
Then 8.2
8.0
But we predicted as much, just maybe not by so much and so soon.
Our second race was tight but we got through and set a personal best of 8.8 seconds, not a world record, but a step closer.
Unfortunately our last race was against an 8.4 world record holder, we did our best, broke a car wheel, had our models not upload to car and raced a very hard to get 9.2 seconds.
We were beaten but not out, we entered the “Last Chance Bracket”.
Unfortunately again we couldn’t make the cut and lost to the last racer we had actually beaten, using the same models/cars and track that we set our world record on. We could only get a time of around 11.4 seconds.
Strange thing about DeepRacer, it might give you straight 8 seconds, then it might seem like an impossible challenge.
A challenge we saw all racers face, including the current world champion at times.
Prepare for the press
After winning the first day, placing top in our group and then following up with a world record run of 8.9 seconds, we started to get some media attention, very quickly.
https://www.youtube.com/watch?v=0y2_t4vV2jk&list=PL0vhBOrwsa9E09tZfIebGzNimSI_5AH0E&
So answer honestly, do your wildcard company proud and get back to racing!
Impressions of the competition
It was difficult at times with anxiety building to wait your turn to race, wifi issues across different tracks slowed progress but everything was ran fairly and evenly across all racers, and that’s what everyone wanted to see, to get a fair chance at the crown for all their hard work.
Win or lose, enjoy yourself and learn how to improve
Losing can seem like a relief when you are 10 hours or longer into your race day but if you can take as many notes and findings as you can it will be your best experience for future victories.
Plans for 2020
Win
To sum up the total experience I have learned a tonne about DeepRacer, what commitment it takes to win and it has really fuelled my interest in ML going forward into the new year.
Thanks for reading!
And as personal extra support I had my amazing fiancée, Fiona Cunningham supporting me at every single step. X
