Read on to find all the results and learn who has become the ultimate AWS DeepRacer League champion.
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.
You might be interested in other posts from our blog
- AWS DeepRacer League Championships 2024 – the last race
- AWS DeepRacer League 2024 Championship Finals – The Last (Competitive) Dance!
- My Journey with DeepRacer
- Revolutionizing Computer Science Education: The DeepRacer Autonomous Race Experience
- DeepRacer-for-Cloud v5.2.2 now available with new real-time training metrics
- Announcing AWS DeepRacer League 2024. New year, new rules
After an exciting and very competitive fight throughout the year, 26 racers from all continents but Antarctica have come to AWS re:Invent in Las Vegas to claim the last AWS DeepRacer League title ever. Throughout the six years AWS DeepRacer has not only helped over 500,000 people on their learning path into machine learning but also kindled (and crushed) dreams, forged friendships, brought joy and non-zero amount of near heart attack states among those who decided to compete. Read on to learn how we did in the subsequent phases of the championships.
If you’d like to see detailed rules that we followed this year, check out Mark’s post on the rules: https://blog.deepracing.io/2024/11/18/aws-deepracer-league-2024-championship-finals-the-last-competitive-dance/.
The Wildcard Race
The group of finalists was to be complemented by six last minute contenders. This year’s format also made the corporate wildcards compete for their spots – winners of internal races from the organisations that have made amazing contributions to DeepRacer’s success through education, competition and contributions to the community.
The race was packed with legends of the DeepRacer scene: 0
- RayG from Singapore – a community member and author of logging tools
- Trackboss – David Smith, former Amazonian and author of the DeepRacer Events Manager (DREM)
- kashi – one of long-running racers and finalists
- Jochem – who competed in DeepRacer from the very first AWS Summit race in Amsterdam in 2019 to the very last wildcard race, but also spent days building and supporting the community
- me, Breadcentric – I’m legendary in a sense that there is a legend that I once could compete with other racers not only to lose
- Duckworth – author or the main contributor into, amongst many others, DeepRacer for Cloud solution, DeepRacer firmware for RaspberryPi, and custom tracks generator
- Jerec – who has been one of the pillars of the community since 2019, and a third racer in 2023 championships
- leadingai – who has dedicated himself to educating and supporting students on their path into technology; one of his mentees, AIDeepRacer, qualified into Round 1
- MattCamp – the only constant on the podium at the AWS Summit in London, one of the first contributors to the DeepRacer for Cloud solution and the community leader
As well as some newcomers: –
- KnightRider – A new competitor to the scene from Eviden, who had their own internal league in 2024 and tried his luck at re:Invent
- SimonHill – A new competitor who won one of the corporate wildcard spots as winner of the Eviden internal competition and was guaranteed entry into the wildcard race
- geervani – A new competitor who won the JPMC female internal league to bag a corporate wildcard spot and was guaranteed entry into the wildcard race
- CyberDrifters – new competitors, another corporate wildcard, this time from Audi and was guaranteed entry into the wildcard race
For those who have experienced racing on the track it is worth mentioning one significant improvement. Both the cars and DREM (DeepRacer Event Manager) have been updated with a mod that performs the model optimisation before its loading for racing. As a result instead of nerve-wracking 30-40 seconds, the model change during the race was taking 2-3 seconds. It still felt like an eternity when my qualification was on the line, but that essentially meant that if I had an off track, by the time the track boss was putting the car back on the track I already had a new model loaded and ready to go. Big thanks to Lars (Duckworth) and David (TheTrackBoss) for preparing the mod, and for the crew for taking the time to update the devices.
Most of the racers have raced only once but we had a few with some more tries on the day. The length of the racing session was just right for everyone who wanted to participate in the race, and for those who got some wonky cars or had more models to compete, and those at risk of falling out, to do a retry. As usual, all of the racing was made special with Ryan Myrehn and Nicole Murray commentary. Throughout the years they have grown as experts not only on the intricacies of DeepRacer as a sport but also the community members, making it extra special for the participants, both new and seasoned in the challenge.
After multiple twists which included stories of broken cars, off tracks, and not-so-well-behaved models, the top six has emerged including the following racers:
- MattCamp representing France
- SimonHill representing Great Britain
- geervani representing India
- leadingai representing Canada
- Jerec representing the USA
- Duckworth representing Germany
These are the final results of the race:
| Rank | Racer | Country | Fastest Lap Time | Average Time | Lap completion ration | Attempts |
|---|---|---|---|---|---|---|
| 1 | MattCamp | FR | 7.791 | 15.466 | 90.0% | 2 |
| 2 | SimonHill | GB | 8.695 | 10.601 | 100.0% | 2 |
| 3 | geervani | IN | 8.757 | 12.526 | 100.0% | 2 |
| 4 | leadingai | CA | 9.527 | 16.591 | 90.0% | 2 |
| 5 | Jerec | US | 9.701 | 13.206 | 90.0% | 2 |
| 6 | Duckworth | DE | 10.027 | 15.308 | 100.0% | 2 |
| 7 | KnightRider | GB | 10.309 | 14.378 | 90.0% | 2 |
| 8 | CyberDrifters | DE | 11.38 | 18.854 | 100.0% | 3 |
| 9 | Breadcentric | PL | 14.281 | 19.914 | 70.0% | 1 |
| 10 | Ayako | JP | 15.164 | 20.053 | 100.0% | 1 |
| 11 | Jochem | NL | 16.241 | 28.101 | 90.0% | 2 |
| 12 | kashi | JP | 16.294 | 31.985 | 100.0% | 2 |
| 13 | Berbbobs | GB | 16.883 | 30.206 | 100.0% | 1 |
| 14 | Trustbase_Kotake | JP | 18.112 | 24.862 | 90.0% | 1 |
| 15 | GroovyDNPds | JP | 22.401 | 29.622 | 70.0% | 1 |
| 16 | Trackboss | GB | 25.814 | 33.188 | 60.0% | 1 |
| 17 | eight | JP | 32.262 | 37.861 | 100.0% | 1 |
| 18 | pankajdgupta | IN | 39.729 | 45.09 | 100.0% | 1 |
| 19 | TrustBase_TANAKA | JP | 40.764 | 48.559 | 100.0% | 1 |
| 20 | RayG | SG | 64.641 | 69.825 | 100.0% | 1 |
Round 1
Round one was split into two groups racing over Monday and Tuesday.
It is important to highlight how professional and well prepared AWS were. While on the outside it may just look like a simple toy with some silly fun event, the winner this year was walking away with a $25,000 cash prize so a lot was at stake. When carrying out the tournament there are many moving parts that require tending to if the races are to highlight the essence of the racing, starting with the proper setup of the venue and training of the pit crew, through enforcing areas seemingly out of the crew’s control such as the Wi-Fi saturation, to preparing a quality live stream and the most important of all, simply caring about the whole thing. Throughout the years we haven’t just been an amazing supportive and inclusive community and a professional team of service providers. AWS DeepRacer has forged friendships that will go beyond the league, something I will be forever grateful for.
To prepare for round 1 each competitor had 2 x 2 minute practice attempts, apart from those who’d took part in the wildcard race who were limited to one practice attempt. As is often the case with DeepRacer the fusion of model and well calibrated car is required to excel and Mark Ross set extremely fast times in practice that sadly (for him!) couldn’t be repeated in round 1.
Round 1 Part 1 on Tuesday comprised of odd numbered wildcard qualifies (1st, 3rd, 5th), Asia Pacific, Greater China and Middle East & Africa. At the end of the day Mattcamp looked secure with an 8.223 average, whereas the rest of those in the qualifying places with average times of 8.985 – 11.769 looked vulnerable to those waiting for their go on Wednesday from the other half of the draw.
Round 1 Part 2 on Wednesday comprised of the even numbered wildcard qualifiers (2nd, 4th, 6th), North America, South America and Europe. It could be argued racers on Wednesday had an advantage as they knew what time was required to qualify. However, as with many things DeepRacer it didn’t quite pan out as expected, with a number of seasoned racers not able to get competitive enough times on the board, perhaps the pressure of this being the final race got to a few people. After having to sweat it out for 24 hours geervani, TonyJ and CodeMachine were pleased to find their times from Tuesday were enough to qualify them for the finals, along with MattCamp who’s time always looked safe and the qualifiers from Wednesday – AIDeepRacer, SimonHill, rosscomp1 and PolishThunder. The events were not without some drama, with the very final spot in the top 8 going to video review, with MarkRoss potentially taking the final spot in the top 8, having been given a third go due to the mechanical issues with the car in his first run that wasn’t restarted by AWS at the time as had happened with some other racers during their runs. Fortunately for PolishThunder, the video review was able to show MarkRoss left the track and his heart rate was able to return to normal levels so he could prepare for the top 8.
As with every sport (and visit to Las Vegas!), if you’re one of the best, come prepared and get your portion of luck you’re in with a great chance, with 50% of the wildcard racers making it to the top 8, against 19% of the pre-qualified 26 racers.
| Rank | Racer | Country | Fastest Lap Time | Fastest Average Time | Lap completion ration | Attempts |
|---|---|---|---|---|---|---|
| 1 | AIDeepRacer | US | 7.85 | 8.102 | 100.0% | 2 |
| 2 | MattCamp | FR | 8.035 | 8.223 | 100.0% | 2 |
| 3 | SimonHill | GB | 7.787 | 8.284 | 100.0% | 2 |
| 4 | rosscomp1 | US | 8.487 | 8.917 | 100.0% | 2 |
| 5 | geervani | IN | 8.54 | 8.985 | 100.0% | 2 |
| 6 | TonyJ | NZ | 8.829 | 9.023 | 100.0% | 2 |
| 7 | CodeMachine | AU | 8.857 | 9.035 | 100.0% | 2 |
| 8 | PolishThunder | US | 8.581 | 9.039 | 100.0% | 2 |
| 9 | JPMC_BA_MiguelG | AR | 8.303 | 9.08 | 100.0% | 2 |
| 10 | leadingai | CA | 8.859 | 9.1 | 100.0% | 2 |
| 11 | FiatLux | LU | 9.118 | 9.223 | 100.0% | 2 |
| 12 | Liao | TW | 9.073 | 9.26 | 100.0% | 2 |
| 13 | MarkRoss | GB | 7.89 | 9.924 | 100.0% | 3 |
| 14 | Duckworth | DE | 9.335 | 10.072 | 100.0% | 2 |
| 15 | Nevertari | EG | 8.47 | 10.113 | 70.0% | 2 |
| 16 | ZoD | CH | 9.596 | 10.188 | 100.0% | 2 |
| 17 | BSi | ID | 10.382 | 11.769 | 100.0% | 2 |
| 18 | XMaroRadoX | EG | 10.086 | 12.145 | 100.0% | 2 |
| 19 | OffTrack | US | 10.85 | 12.152 | 100.0% | 2 |
| 20 | JJ | CA | 9.087 | 12.168 | 100.0% | 2 |
| 21 | Jerec | US | 10.781 | 12.404 | 90.0% | 2 |
| 22 | Nariox | CH | 10.63 | 13.054 | 100.0% | 2 |
| 23 | JPMCBAD3 | AR | 10.752 | 13.113 | 100.0% | 2 |
| 24 | KamilojADL | CO | 8.598 | 13.29 | 100.0% | 2 |
| 25 | yinwun | CN | 10.527 | 13.358 | 100.0% | 2 |
| 26 | mkreder | AR | 13.174 | 14.038 | 100.0% | 2 |
| 27 | HArchitect | CN | 10.324 | 14.143 | 100.0% | 2 |
| 28 | DBro | GB | 11.708 | 14.155 | 100.0% | 2 |
| 29 | BennyLBS | CN | 13.792 | 15.199 | 100.0% | 2 |
| 30 | dpkkdn | IN | 13.516 | 17.572 | 100.0% | 2 |
| 31 | ryu | GH | 22.657 | 23.524 | 80.0% | 1 |
| 32 | yuliya | KZ | 25.913 | 31.149 | 100.0% | 2 |
Round 2
Our top 8 finalists moved on to the head to head racing format in round 2, with seeding between them meaning AIDeepRacer in 1st place facing off against PolishThunder in 8th place, 2nd facing 7th etc.
Prior to round 2 on the evenings after round 1 each qualified racer was given 1 x 2 minute practice attempt on the track, as round 2 was being raced in a counterclockwise direction, making the racing a completely different challenge to round 1 in the clockwise direction. Having had their practice attempts on Tuesday or Wednesday evening some racers went back to their rooms to train new models, such was their concern about what they’d seen in practice. This gave a small advantage to those sweating on a place after Tuesday’s round 1, with a whole 24 hours more than Wednesday’s racers to train new models if they so wished, although they wouldn’t be able to see the fruits of their efforts until competitive racing started on Thursday. The head to head racing was based on up to 2 x 2 minute runs each, with winners progressing in the undefeated bracket and losers dropping into the defeated bracket.
PolishThunder, having an impeccable record of racing on a Thursday as previous re:Invents (when the main competition was over and people were competing to prequalify for the next year) seemed in buoyant mood, however AIDeepRacer showed his top seed credentials, quickly booting PolishThunder into the last chance bracket.
New face on the scene, SimonHill, quickly showed the wildcard card race and round 1 was no fluke, dispatching with TonyJ and MattCamp with only a single 2 minute attempt, as neither could beat his time and force a rebuttal.
Some great friends from the community ended up having to face each other during the top 8, with PolishThunder and rosscomp1, as well as MattCamp and TonyJ having to temporarily suspend their friendships for a few minutes to battle each other on the track.
After hours of racing AIDeepRacer and SimonHill in the undefeated bracket were joined by PolishThunder, who held his nerve having lost in the very first race in the round to battle his way through to the final 3.
Interestingly, as you can see from the variance in times achieved in the brackets, some of the cars used during the races were significantly slower than others, demonstrating that car calibration can have a significant impact on times achieved. It’s testament to the racers having models that were able to put together a 3 lap average when facing sub-optimal conditions, with some models being extremely good at tolerating the wide variance of car calibration.
The finale
Prior to the final 3 starting there were some additional steps for the racers to go through. To ensure fairness and an inability of each racer to impact another’s chances by crashing the cars each racer had to have their own car, which they were free to recalibrate, either with the assistance of the AWS Pit Crew or community members.
Two cars had proved to be reliable in the round of 8, and PolishThunder, having set the fastest time in match 9, and AIDeepRacer, having set the 2nd fastest time in match 1 go to choose their cars first and they chose the two cars that had ran well. This left SimonHill with a dilemma as to which car to choose, with him eventually settling on running ‘Expo2’, which had set the lap record times in a clockwise direction in round 1 free practice, prior to being calibrated and not having been as good the rest of the week.
PolishThunder worked with the AWS pitcrew to configure his car. AIDeepRacer and SimonHill worked with community members to configure their cars, some utilizing ‘calibration’ models that were only cable of performing a single action (0 degrees, 30 degrees right or 30 degrees left), and cars were seen doing donuts in the cage to make sure their calibration was optimal.
The direction of the final 3 race was important, PolishThunder having set the fastest average time in a counterclockwise direction in round 2, but only having scraped into the top 8 with his clockwise model. There was a divergence of strategy between racers, with PolishThunder having separate models for each direction, whilst SimonHill was using the same model regardless of direction. Racers wondered what their fate would be as the coin went up and came down on counterclockwise for the final competitive race in AWS DeepRacer history…
Each racer was given up to 1 minute to check they were happy with their car and its calibration. PolishThunder was that confident at the result of the coin toss he only did a single practice lap and stopped the car, it was looking very good. SimonHill did a few practice laps and also looked strong. Sadly for AIDeepRacer, who had put in a strong performance all week, his car kept going off at the 2nd last corner but no objections were raised to change cars.
Polish Thunder went first, having qualified via the defeated bracket. He set a very strong 3 consecutive lap average time of 8.436s under the pressure of having to go first and set the bar as high as he could for others to follow.
SimonHill went next, and having dialed in his model was able to set the fastest 3 lap counterclockwise time achieved all week at 8.353s, knocking PolishThunder down to 2nd place with lest than 0.1s separating their times. It wasn’t to be for AIDeepRacer with his car performing in a similar way to his 1 minute test, but he proved some great skill altering his relative speed of the car to get a 3 lap average of 11.246s on the board.
And with that we had our 2nd consecutive corporate wildcard winner, with Eviden’s SimonHill taking over from Accenture’s FiatLux as the AWS DeepRacer Champion!
All three finalists, along with MattCamp, rosscomp1 and TonyJ rounding out the top 6 were presented with their prizes (cash, DeepRacer champions jackets, trophies, depending on their final position). ZoD, as winner of the virtual league throughout the season joined them. The AWS DeepRacer family, comprising of the AWS pit crew, racers and community members met on the track one last time to take some pictures together, there was definitely a mixture of joy at having the opportunity to get together one last time, as well as sadness that the competitive league was coming to an end.
Simon being fitted for his jacket (source – KnightRider) 
Simon with the winner’s trophy and jacket (source – KnightRider) 
Simon being interviewed by Ryan (source – KnightRider) 
Simon with the car trophy made by an Formula 1 trophy maker (source – Jochem)
Along with the normal prizes on offer to the winner SimonHill was later presented with a silver DeepRacer trophy, made by one of the people who makes trophies for Formula 1! Simon was a very humble winner, commenting on how the community was so welcoming and had helped him along his DeepRacer journey, although he did also have some fun with the trophies!
Lionel Messi after 2022 World Cup (source – Lionel Messi’s Instagram) 
Simon Hill – our 2024, and final, Champion (source – Simon Hill – yes you can see the selfie reflection on the trophy 🙂 )
A few tips from our champion!
We caught up with Simon about how he went about taking victory at the final competitive event.
Simon was one of the few racers to utilize the same model for both clockwise and anti-clockwise racing, so how did he do it?
Simon reports he created a well generalized model by training on multiple tracks. He utilized 6 different tracks and trained on them in both directions. This meant his model was presented with a large variety of images during training, although non of them had a giant GitLab sign on the ceiling as models faced in the expo! Simon thinks this was key both to generalization as well as prevented him from having an over fit model, despite training for ~24 hours. Simon selected tracks that were the same width as the Forever Raceway (~76cm) and had similar corner characteristics as he expected to face on the Forever Raceway, along with varying backgrounds to avoid over fitting to the plain background of some of the tracks.
Simon’s believes his action space was key to his success, as it created a forgiving model that was able to slow down when it was in less desirable areas of the track. Some other competitors models would just head off into the wall when faced with similar track positions, and this could have made Simon’s model more tolerant to sub optimally calibrated cars, as it was the most consistent model throughout the week. Simon achieved this by having multiple speeds per steering angle, allowing the model to maintain course whilst adjusting speed as appropriate. This may come as a surprise to seasoned racers, whom the gospel has been to train at lower speeds for a number of years, but Simon’s action space included 4m/s actions.
Simon developed a reward function that was able to avoid weaving from side to side, which also looked important given the amount of straight sections on the track.
Data Driven Racing via DREM
Enjoyment of those attending in person as well as those watching online via Twitch has been greatly improved in recent years with the excellent AWS DeepRacer Event Manager (DREM). It provides all of the lap times via a pressure sensor on the start / finish line for accuracy, along with the streaming overlay so people can see lap times, time remaining etc.
Thanks to DREM we have access to all the leaderboards from the week, allowing us to pull out information like lap records. We can even get both the fastest single lap and average laps, regardless of the format of the race.
Using ths data we can see the following lap records for the ‘Forever Raceway’ were set: –
| Track | Lap Type | Racer | Time | Race |
| Forever Raceway Clockwise | Single Fastest | MarkRoss | 7.574 | Practice Round 1, Part 2 |
| Forever Raceway Clockwise | 3 Lap Average | MarkRoss | 7.744 | Practice Round 1, Part 2 |
| Forever Raceway Counter Clockwise | Single Fastest | AIDeepRacer | 8.142 | Round 2 – Bracket of 8 |
| Forever Raceway Counter Clockwise | 3 Lap Average | SimonHill | 8.353 | Round 3 – Finale |
Links to all the leaderboards, both for fastest single lap and average lap, for all sessions at re:Invent 2024 are provided below: –
Wildcard (Monday): –
Average – https://dhwxpgndrwxhz.cloudfront.net/leaderboard/c7cfd83c-df5a-4548-9418-24ef972787f5/?scroll=false&track=1&format=average
Practice Round 1 Part 1 (Tuesday): –
Average – https://dhwxpgndrwxhz.cloudfront.net/leaderboard/be6ee697-c648-45c0-8125-864de7680a60/?qr=header&scroll=true&track=1&format=average
Practice Round 1 Part 2 (Wednesday): –
Round 1 – Combined
Round 2 – Bracket of 8
Round 3 – Finale
DeepRacer Forever!
With the conclusion of the AWS competitively sponsored racing the virtual league and racing at summits and re:Invent have ended, so what’s next?
From a DeepRacer perspective it will live on in 2025 within the AWS Console, before being released as an AWS solution that can be deployed into your own AWS account. This will allow hobbyists and corporations to continue to use DeepRacer to continue to drive learning in machine learning, data analysis and with workshops added to DeepRacer in the last two years, Generative AI. AWS have tracks and will continue to run some corporate events in 2025, and companies like Eviden, JPMC and Accenture have tracks that they may continue to use.
From a wider perspective AWS trialed the finals of the ‘League of LLMs’ at re:Invent, building on what has taken place this year in the APJ region. It was great to see community member RayG involved in that. Will we see this become more mainstream in 2025, I guess only time will tell…
To keep up to date on all things DeepRacer it’s best to stay informed via: –
GitHub Community Organization – https://github.com/aws-deepracer-community
Join us on Discord – https://join.deepracing.io/
Blog credits – written by Tomasz Ptak and Mark Ross