Only one week of regular races left in AWS DeepRacer League 2021. While we wait for the full announcements on this year’s championship format, let’s have a look at where we are on the track.
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
September finale
Be sure to catch up on the September finale race update.
Championships announcement
AWS have announced that the 2021 Championships will happen virtually only. Because of this 2020 and 2021 trips have been replaced with value payouts.
Unfortunately this means we will not get to meet in person at AWS re:Invent this year.
NeurIPS 2021 DeepRacer Challenge
There’s a totally new challenge available and you can join it in October. Winners will get AWS credits, DeepRacers and ability to co-author a paper about their rewarding strategy.
More details: https://blog.deepracing.io/2021/09/16/invitation-neurips-2021-aws-deepracer-challenge/
Udacity scholarship program results are out
Udacity have released results for the AWS Machine Learning Scholarship Program. Tens of thousands of students registered for the program before its start in June, we’ve had over 36,000 new members join our Community Slack and support each other on the learning journey.
This week Udacity have sent out decisions on who makes it into the nanodegree course in Phase 2. If you are one of the students and have not received any emails (or are trying to find your certificate of completion), do read recent communications in our Slack’s #general channel.
Pro Division
Race info:
- 69.96 meters long
- 107 centimeters wide
- Head-to-bot format with four bot cars driving at a mental 4 m/s with no lane changes
- Racers are trying to complete three laps – all laps are separate and from a standing start
- Each collision results in a five seconds penalty
- Each driving off the track results in a three seconds penalty
- Top 16 go on into a live race which will be streamed on Twitch to fight for places in the League Finals
- 10 best racers who haven’t yet done so win an AWS DeepRacer EVO (last one in September was in 42nd place if I got the numbers right)
All submissions Update time: 2021-10-24 13:00:00 UTC Racers: 204 Total submission attempts: 44826 Best total time: 01:07.276 by JJ Mean total time: 03:55.825406863 Best lap: 00:22.319 by JJ Mean best lap: 01:11.478127451 Mean average lap: 01:18.608137255
While the leader is not at all surprising, I would like to draw your attention to the time. You must admit that JJ’s times are a measure of the tracks’ complexity and 01:07 is his second slowest time this season. Even more remarkable is the difference in the mean total time which is almost 90 seconds slower than the the corresponding value from September and has so far been the slowest mean time to complete the track in 2021’s Pro Division. This just shows how much difficulty Expedition Super Loop presents.
204 racers is a stable number of Pro racers that we’ve seen since August when Udacity students started qualifying from the Open division and Pro participants count jumped from 120-130 to 200-210. We may expect a few more to show up in the last week.
Top 25 submissions Update time: 2021-10-24 13:00:00 UTC Racers: 25 Total submission attempts: 25589 Best total time: 01:07.276 by JJ Mean total time: 01:22.176200 Best lap: 00:22.319 by JJ Mean best lap: 00:26.222080 Mean average lap: 00:27.391720
Again, mean total time for the leading group is amongst the highest ones in 2021. Also the difference between the mean total time and the best total time is about 15 seconds compared to 7-11 seconds in the previous five races. I can see two potential reasons, one being the difficulty of the track and the second being general drop in activity this month – we have observed submission counts halve compared to September. That said we’ve had between one and four resets in top group submissions in previous months and this time it’s twenty which makes me lean towards the challenging aspect of the race.
Let’s have a look at the leaders:
| Rank | Racer | Total Time | Avg Lap | Best Lap | Resets | Submissions |
|---|---|---|---|---|---|---|
| 1 | JJ | 01:07.276 | 00:22.425 | 00:22.319 | 0 | 4352 |
| 2 | Karl-NAB | 01:08.885 | 00:22.961 | 00:22.713 | 0 | 658 |
| 3 | rosscomp1 | 01:12.554 | 00:24.184 | 00:23.862 | 0 | 595 |
| 4 | JPMC-Rogue-Hyderabad | 01:12.702 | 00:24.234 | 00:23.210 | 0 | 149 |
| 5 | Duckworth | 01:12.986 | 00:24.328 | 00:24.075 | 0 | 3082 |
| 6 | Ernesto | 01:13.643 | 00:24.547 | 00:23.948 | 0 | 1852 |
| 7 | Yi-LI-NYCU-CGI | 01:14.379 | 00:24.793 | 00:24.317 | 1 | 197 |
| 8 | ZhengYi-NYCU-CGI | 01:15.966 | 00:25.322 | 00:25.060 | 0 | 345 |
| 9 | CGS-BigBoson | 01:18.855 | 00:26.285 | 00:25.857 | 0 | 5047 |
| 10 | ADA | 01:19.942 | 00:26.647 | 00:25.917 | 0 | 6 |
| 11 | Penguin-DNPds | 01:20.148 | 00:26.716 | 00:24.866 | 1 | 498 |
| 12 | Matt | 01:21.140 | 00:27.046 | 00:26.409 | 0 | 1052 |
| 13 | JPMC-HYD-CircuitBreakers | 01:21.196 | 00:27.065 | 00:26.592 | 0 | 457 |
| 14 | CGS-LargeLattice | 01:22.288 | 00:27.429 | 00:27.164 | 0 | 928 |
| 15 | CGS-Huge-Hadr0n | 01:22.373 | 00:27.457 | 00:25.524 | 2 | 799 |
| 16 | TonyJ | 01:23.388 | 00:27.796 | 00:25.160 | 2 | 286 |
| 17 | JPMC-Singapore-Chetak | 01:23.626 | 00:27.875 | 00:27.076 | 0 | 101 |
| 18 | Deepak-dpk | 01:23.793 | 00:27.931 | 00:27.480 | 0 | 1063 |
| 19 | AJM-Model-1 | 01:24.186 | 00:28.062 | 00:26.001 | 2 | 17 |
| 20 | Fumiaki | 01:26.933 | 00:28.977 | 00:24.743 | 4 | 1799 |
| 21 | GT-DevelopersIO | 01:27.164 | 00:29.054 | 00:25.931 | 2 | 53 |
| 22 | PolishThunder | 01:30.280 | 00:30.093 | 00:27.200 | 1 | 4 |
| 23 | Bandwidth-RnD | 01:40.949 | 00:33.649 | 00:33.369 | 0 | 827 |
| 24 | Aurora-DNPds | 01:44.222 | 00:34.740 | 00:33.757 | 1 | 1106 |
| 25 | CGS-on-your-12-hopefully | 01:45.531 | 00:35.177 | 00:33.002 | 4 | 316 |
The top 16 finale contenders barely fit into the top 25 which means that many of the championship finalists keep racing, perhaps as a warm-up before the finals?
Compliments go to ADA, AJM-Model-1 and PolishThunder for making it into and staying in the top 25 with fewer than 20 submissions so far. This stands out for PolishThunder especially as he used to be one of the top submitters month after month.
The 2019 runner-up Fumiaki is still in the race which means we might have as many as three DNP racers in the finale together with Pinguin-DNPds and Aurora-DNPds.
Open Division
Race info:
- 44.04 meters long
- 107 centimetres wide
- Simpler and shorter than the pro track but still pretty challenging with a variety of straights and turns
- Time Trial format
- Competitors need to complete three laps and total time counts for their classification, each lap starts from halt
- Each off-track results in three seconds penalty.
- Racers fight for qualification to the Pro group which is won by the top 10%
- They will also receive a Pro Welcoming Kit which is a bit of a mystery (hoodie, cap and non-slip socks are what we’ve seen so far)
I’m actually curious about the effect of winning the Open division – will there be a last minute chance to fight for spots in the finals? Will there be a Pro Division next year? I reckon we’ll find out soon but now let’s have a look at what our racers are up to.
All submissions Update time: 2021-10-24 13:00:00 UTC Racers: 2196 Total submission attempts: 13261 Best total time: 01:16.828 by DoomBuggy2 Mean total time: 03:35.531479053 Best lap: 00:23.935 by DoomBuggy2 Mean best lap: 01:07.196897086 Mean average lap: 01:11.843489071
The number of racers has increased significantly compared to 1941 overall racers in September or 1635 in August and is only lower than the July count of 2433. I suspect the July race was mostly participated by the early starters of the scholarship’s Foundations course and October is primarily attended by last minute challengers.
Open track is significantly simpler than the Pro one yet most racers are ahead of their greatest learnings so it is not surprising that best times are worse than those in the Pro division. DoomBuggy2 seems to be keeping things under control. I suspect the third DoomBuggy is the same person but using a different account but I’ll leave that for the AWS to clarify as needed.
Top 25 submissions Update time: 2021-10-24 13:00:00 UTC Racers: 25 Total submission attempts: 647 Best total time: 01:16.828 by DoomBuggy2 Mean total time: 01:45.118440 Best lap: 00:23.935 by DoomBuggy2 Mean best lap: 00:32.964 Mean average lap: 00:35.039160
There’s quite a big gap between the results for the whole group and the top 25. This is expected in Open Division which often also contains accidental submissions from the newest racers after an hour or two of training on some basic track. That’s how mean total time in top 25 is two times shorter than the one for all racers.
| Rank | Racer | Total Time | Avg Lap | Best Lap | Resets | Submissions |
|---|---|---|---|---|---|---|
| 1 | DoomBuggy2 | 01:16.828 | 00:25.609 | 00:23.935 | 1 | 9 |
| 2 | ovbiedo-pataki | 01:20.251 | 00:26.750 | 00:26.315 | 0 | 375 |
| 3 | DoomBuggy | 01:21.867 | 00:27.289 | 00:26.006 | 1 | 12 |
| 4 | Racer0 | 01:29.062 | 00:29.687 | 00:26.800 | 8 | 2 |
| 5 | awsandy | 01:30.744 | 00:30.248 | 00:28.537 | 1 | 19 |
| 6 | BeepBeepRibbyRibby | 01:40.056 | 00:33.352 | 00:32.724 | 3 | 18 |
| 7 | Geronimo | 01:40.375 | 00:33.458 | 00:31.665 | 1 | 2 |
| 8 | JPMC-BMTH-DarkStars | 01:42.061 | 00:34.020 | 00:31.862 | 5 | 17 |
| 9 | BrainyTest#7599 | 01:42.525 | 00:34.175 | 00:31.531 | 5 | 6 |
| 10 | NguyenTrungThanh | 01:46.473 | 00:35.491 | 00:30.941 | 12 | 6 |
| 11 | Sigmoid | 01:46.819 | 00:35.606 | 00:33.277 | 1 | 45 |
| 12 | JPMC-HongKong-S-paceX | 01:46.953 | 00:35.651 | 00:34.625 | 12 | 32 |
| 13 | RomyRacing | 01:49.846 | 00:36.615 | 00:34.209 | 8 | 15 |
| 14 | BetaPatch#1919 | 01:51.053 | 00:37.017 | 00:36.200 | 12 | 2 |
| 15 | Pangzi | 01:51.182 | 00:37.060 | 00:33.934 | 14 | 12 |
| 16 | AutoSpeedingBalancer | 01:52.589 | 00:37.529 | 00:34.984 | 5 | 7 |
| 17 | DorothyChen | 01:52.997 | 00:37.665 | 00:32.928 | 16 | 8 |
| 18 | JaguarRobotics-Blaze | 01:53.326 | 00:37.775 | 00:34.754 | 2 | 12 |
| 19 | Unsupervised-Racers | 01:54.317 | 00:38.105 | 00:36.738 | 7 | 2 |
| 20 | Ainat | 01:54.653 | 00:38.217 | 00:37.463 | 3 | 18 |
| 21 | capstone-018-addspeedweight-3 | 01:55.343 | 00:38.447 | 00:36.868 | 14 | 3 |
| 22 | JPMC-MUMBAI-FormulaAI | 01:55.990 | 00:38.663 | 00:36.597 | 15 | 3 |
| 23 | 5thStWaffleHouse | 01:56.941 | 00:38.980 | 00:37.212 | 9 | 9 |
| 24 | RushingBear-NUK | 01:57.717 | 00:39.239 | 00:35.994 | 9 | 8 |
| 25 | PrimeServer#4161 | 01:57.993 | 00:39.331 | 00:38.001 | 14 | 5 |
What draws my attention in the Top 25 list is that there is only one racer with zero resets: ovbiedo-pataki. Other runs seem to be having difficulties mainly around the crazy chicane section, as if their models were unable to drive past the turns – could it be problems with the action space being too fast for this section? I’ve especially looked at Racer0 and it seems like it, the off-tracks are happening only in those turns. Alternatively the models might not have been trained enough but I don’t think it’s that.
Next week we will learn who enters the last planned finale race for the 2021 season of the AWS DeepRacer League. Soon we should also learn what the championships will bring. Remeber to visit for our report from the tracks.