I’m very excited to share with you the report on AWS DeepRacer League 2021 March Pro Finale
It’s the first time that we could watch this type of race. In the 2021 season we have two divisions: Open where racers fight to qualify into the Pro division, and the Pro division where racers try to get into the finale race and to fight for a place in the great final which (conditions permitting) will take place in Las Vegas during the AWS re:Invent conference. Each month three tickets are up for grabs! Also, ten best racers win a DeepRacer Evo.
There are two rounds of runs. In a round each racer has a 3 minutes time slot to clock the fastest possible lap. The order in round one is based on results from the full month’s race (slowest to fastest), and in round two it’s based on the results in round one. The final result for each racer is the best lap from both sessions.
The race format is object avoidance with four boxes on track and off-track/collision penalties at 5 seconds each.
You can watch the whole three hours of stream on Twitch:
But if you prefer to just go through a summary, keep reading.
AWS DeepRacer
AWS DeepRacer is a 1/18th scale autonomous race car but also much more. It is a complete program that has helped thousands of employees in numerous organizations begin their educational journey into machine learning through fun and rivalry.
Visit AWS DeepRacer page to learn more about how it can help you and your organization begin and progress the journey towards machine learning.
Join the AWS Machine Learning Community to talk to people who have used DeepRacer in their learning experience.
DeepRacer News
Create your customised racer profile and tweak your car’s looks
AWS’ Joe Fontaine has shared new features of the AWS DeepRacer Console in an AWS blog post.
First in-person meetup of 2021
If you’re from Indiana or generally close to Fort Wayne, do join the AWS Hero Alex Schultz for the first in-person DeepRacer meetup of the year on the 28th of April. Read more in a separate blog post.
First Round
The race started with Joffe whose car sadly set its mind on one box so much it failed to get past it and complete the lap. Colliding with an object on track results in a five seconds penalty and a reset of the car slightly back from its most recent position but for some models that may be not enough to get around the obstacle.
Maikel from Argentina is a bit of a mystery driver to the Community as this is the only new-starter in the top 16 which hasn’t made an appearance in our Slack group. This wasn’t a smooth ride with a few resets happening but a solid time of 38.589 s has put Maikel at the top.
JPMC-DriftKing from USA followed. He’s got some experience with this racing format as it has been used in round one of the AWS DeepRacer League 2020 and he qualified into the next round and eventually also the grand prix final. While a slightly slower time of 41.887 s might not sound very impressive compared to Maikel, I’m sure we’ll hear more of him in the future races. In this session he got overtaken straight away by another 2020 finalists Robin-Castro with time of 40.925 s. Dartjason who followed had big problems on another object and a lap time of 2:35.845 was not enough for a chance in the finals.
Ernesto who has been participating in DeepRacer League since 2019 (he qualified into the finals through a last chance race at Aria during the re:Invent) had a pretty steady model overall but just not fast enough to counter the resets that were sneaking in in each lap he made. He tried to address that by increasing the top speed (racers have access to a speed multiplier slider) but that is a tricky thing – at the speeds that they are using cars tend to skid and fall off the track, and there have been a few of those in this run and 44.158 s have put him in the fourth. GT-DevelopersIO from Japan have quickly set a solid lap at 39.540 s but struggled to improve later on.
The next two competitors are veterans – they started racing in the first DeepRacer physical races during AWS Summits of 2019 in Australia and Netherlands. Jochem from Netherlands started with a slower time in the 50 seconds range but his third flying lap had no resets and with 36.937 s put him on top of the table for now. Karl-NAB from Australia started really strong and two clean laps gave him a 32.232 s time and a lead. While he struggled later on and hit a few boxes he was the first racer to complete four laps in a run.
Next was the 2020 finale runner-up, Duckworth has had some box avoidance issues which meant that his time of 47.783 s put him in the eighth, putting all his hopes on a clean run in round two. PolishThunder seems to have set his speed multiplier too high as his delivery truck seemed to be racing on ice. Multiple spins and resets meant that his best time of 52.021 s was simply not enough to get him into the top three. Fumiaki from Japan has been the runner-up of 2019. His car struggled with multiple resets and his best time of 59.509 s means the final hopefuls remained unchanged.
DBro of Northern Ireland started with a few resets and just couldn’t get rid of them in his first run and 46.259 s but him in the eighth, way too slow. RogerRabit has shown some nice manoeuvrers but the resets meant the best time of 40.833 s was too slow for the top three.
The second last two competitors are a bit of a family challenge. flatearth has started with a lot of struggles trying to get past the third box on track. Eventually he managed to get through and spin a very solid lap of 34.899 which put him into second. His father JJ followed. He is a well established name in the world of DeepRacer. He had a pretty solid run and with 34.285 s jumped into second, knocking Jochem out of the group of three.
Let’s see the standings after round one:
| Racer | Round one | |
|---|---|---|
| 1. | Karl-NAB | 32.232 |
| 2. | JJ | 34.285 |
| 3. | flatearth | 34.899 |
| 4. | Jochem | 36.937 |
| 5. | Maikel | 38.589 |
| 6. | GT-DevelopersIO | 39.540 |
| 7. | RogerRabbit | 40.833 |
| 8. | RobinCastro-DBS | 40.925 |
| 9. | JPMC-DriftKing | 41.887 |
| 10. | Ernesto | 44.158 |
| 11. | DBro | 46.259 |
| 12. | Duckworth | 47.783 |
| 13. | PolishThunder | 52.021 |
| 14. | Fumiaki | 59.509 |
| 15. | dartjason | 2:35.845 |
| 16. | Joffe | DNF |
Second round
In this round racers start according to their places so far. This doesn’t change much for Joffe who qualified from the 16th spot. Pretty consistently, his model settled down for box three and ended the run without a complete lap. Dartjason has managed to improve greatly getting down to 50.504 in his first lap. Second lap started rather slow and soon switched to loss of stability and some skidding, followed by an attempt to finish Joffe’s effort. Box three is one tough cardboard, I would’ve given up after such a beating. Dartjason has improved his time significantly and while not good enough to win a place in the finals in Vegas, there is a chance of getting slightly higher in the ranks.
Just a reminder here that the racers don’t just fight for the top 3 placement which puts them in the AWS DeepRacer League 2021 Final with a cost-covered trip to AWS re:Invent conference in Las Vegas. The ten best racers will also receive their own AWS DeepRacer Evo – the physical 1/18th scale car with two cameras and a LIDAR sensor which can have the model loaded, be put on a track and race.
Fumiaki returned to the track with a pretty fast model which has spectacularly smashed into box one. First lap of 1:10.683 was not fast enough to climb up in the standings but he did improve slightly in lap two with 57.055. From the looks of it his model was speeding on the edge, sadly more often on the wrong side of it with a significant amount of off-track resets. In the last attempt he showed off with a superb pass of a box on the edge of the road. It is possible and as you can see some models can do it, usually those trained with more consuming methods that reward the outcomes instead of telling the car what to do. I’m pretty sure this is not the last finale for Fumiaki this year.
PolishThunder returned to the track with a first lap of 40.722 like it’s no biggie. It looked even better in the second loop where he managed to avoid any resets and with 36.618 he joined the fight for higher places. While I wouldn’t let him steer his acrobatics plane with reinforcement learning-trained model just yet it was certainly looking much better. He definitely decided to add some speed which removed the stability and he finished his run with worse times but 36.618 s got him up to fourth for now. Not for long though.
Talk about commentator’s curse: Blaine and Ryan just mentioned that one needs a one fast lap to make it when DBro entered the stage. First lap of 39.894 helped him improve (but not enough), in second he struggled with a bit of wobble in the second sector and with a reset he clocked a time 42.6s. Lap three: best time so far in sector one, best time so far in sector two and best time so far in sector three gave him 33.338 s and DBro entered the fight for the flight. Blaine: if I ever qualify to top 16, please do talk about a risk of doing well when I drive!
Ernesto knows a thing about getting into the finals last-minute – finishing second at the Aria track on the first day of AWS re:Invent 2019 he got himself the entry to the finals the following day. A solid run without resets gave him an opening lap time of 39.711 s. Second lap was slightly slower with 44.7 s – exactly five seconds slower due to a reset. 43 s in lap three followed by the time running out and he has given himself a chance to fight for an Evo in 9th spot.
JPMC-DriftKing presented times of 1:10.066, then 57.452 and since it hasn’t improved, his best time remained at 41.887. I suspect the car had a boosted speed and thus lacked the grip.
RobinCastro-DBS was the first racer to hit all four boxes in a single lap and one of them even twice, sadly no bonus points for this, only penalty seconds. He clocked 1:27.244 in the only lap of the run and remained on 40.833 and by this time was out of the top 10 which means he’ll need to fight a little more to add an Evo to his collection.
RogerRabbit has also put his bet on speed which has lead to a stability decline. First lap of 51.569 was followed by a slightly faster but less stable run of 56.3 seconds and the time has run out for him leaving him with the best time of 40.833
GT-DevelopersIO has started pretty fast but got denied by two resets with 47.927 lap. That improved significantly in the second spin with 40.851 – not great, not terrible. Then 46.2 s. The speed is there but stability is lacking to just push it closer to the podium. GT-DevelopersIO completed with the best time of 39.540.
Maikel opened up with 1:09.986 which is a significantly worse result compared to round one. Second lap of 51.547 was his last and Maikel has completed his start in eighth with 38.589s.
Jochem had to shave almost three seconds from his best lap. Had a pretty strong start at 37.337. If it was a sum of best laps from each round Jochem would definitely still be in the game but that’s not how the finale is done. All other laps sadly suffered from the resets. He only had 4 resets in the whole run but in his situation every penalty just denies a chance for that dream-come-true top three.
Flatearth had lost his podium spot to DBro and began with 46.436 which was nowhere close. Even 39.052 wouldn’t make it – I mean the speed was there: if you remove the reset time it would’ve been enough (even less since there would be no start from zero speed) and it would’ve pushed JJ down. A few more resets scrapped his chances for another try and Flatearth with time of 34.899 landed right behind the podium, sparing his dad’s ticket.
Now it was just about who would grab which place of the podium. JJ jumped in front of DBro in his first lap with 33.334. Second lap much slower at 50.1 s but he did improve in third with 32.864. Fourth ended at over 33 seconds and JJ got himself into the second, giving Karl-NAB a victory run to close the session.
And Karl’s model is just insane. He opened the run with a spectacular 33.084 and then improved even further with 31.151 s in the second lap. Then he got a rather poor 31.7 s and 31.2 s. In the fifth lap (yes, Karl did five full laps) he started by beating the first sector time of the finale but then a reset took his chance to improve even further. Three out of his five laps in this run were enough to win the whole session: that speaks volumes of his skills.
Let’s have a look at the end results:
| Racer | Round one | Round two | Final | |
|---|---|---|---|---|
| 1. (↔0) | Karl-NAB | 32.232 | 31.151 | 31.151 |
| 2. (↔0) | JJ | 34.285 | 32.864 | 32.864 |
| 3. (⬆8) | DBro | 46.259 | 33.338 | 33.338 |
| 4. (⬇1) | flatearth | 34.899 | 39.052 | 34.899 |
| 5. (⬆8) | PolishThunder | 52.021 | 36.618 | 36.618 |
| 6. (⬇2) | Jochem | 36.937 | 37.337 | 36.937 |
| 7. (⬇2) | Maikel | 38.589 | 51.547 | 38.589 |
| 8. (⬇2) | GT-DevelopersIO | 39.540 | 40.851 | 39.540 |
| 9. (⬆1) | Ernesto | 44.158 | 39.711 | 39.711 |
| 10. (⬇3) | RogerRabbit | 40.833 | 51.569 | 40.833 |
| 11. (⬇3) | RobinCastro-DBS | 40.925 | 1:27.244 | 40.925 |
| 12. (⬇3) | JPMC-DriftKing | 41.887 | 57.452 | 41.887 |
| 13. (⬇1) | Duckworth | 47.783 | DNS | 47.783 |
| 14. (⬆1) | dartjason | 2:35.845 | 50.504 | 50.504 |
| 15. (⬇1) | Fumiaki | 59.509 | 57.055 | 57.055 |
| 16. (↔0) | Joffe | DNF | DNF | DNF |
Pretty eventful race, I must admit. Congratulations to Karl-NAB, JJ and DBro who are the first ones to qualify into the finals, and what a group this is!
Make sure you visit again this week for a regular weekly update on the April race.