After attending a AWS summit where we saw the first AWS DeepRacer race, we decided we wanted one at CBM!
So the next day or almost, on March 22nd, we ordered a DeepRacer on Amazon. So exciting! We wanted to organize a race during our annual seminar planned in May.
We were told the very first model would be delivered in April 2019, then delivery was delayed to July. Then no more date. OMG!
So we took matters in our own hands and asked our AWS sales representative to find us a car ready for our end-of-year party, because the seminar was obviously not an option anymore.
We had to be patient, because our sales representative had to activate her internal AWS network in order to find a car, and finally the “Holy Grail” arrived …. Drum roll: early December and our party was to takes place on December 19th.
We established a few rules and guidelines to follow:
- Train in the DeepRacer simulation environment
- Build teams with 3 people including 2 OPS/DEV max per team and one non-technical guy/girl (PO, Designer, Sales…)
- We added a “PimpMyCar” challenge in which each team had to customize the car and then remove it for the next team.
- Each team would have 6 minutes to perform a full round.
And then, the teams started to pour in : we had about 30 people interested among the 70 of IT to take part of this race. We had super catchy names like “Fast & Furious” (yes the name is not patented, right?…) and also more frenchy names like “VroomVroomRacing”. In the end, we had officially 6 teams entered to compete at the CBM Race, a great challenge in sight!
Let the game begins!
A few of the teams had the chance to test on a real track and to win a magnificent price! (Well ok not so beautiful, hey, but it’s still pretty cool to have the honor of finishing first and to show a better time than the the first price at AWS DeepRacer France (12.02), so not bad at all!)
In November, two teams had the chance to race on real physical track during an Axel Springer Group convention.
It was finally time for the big CBM challenge planned on December 19! Only the best 3 teams remained! Yes, projects, vacations and the end of the year had decimated the other teams who did not have sufficient time to work on their model: they declared forfeit.
On December 12th, we rehearsed on the track before D-day to set it up! I can assure you that it is not easy-peasy to set that thing up! You have to like doing puzzles and crafts activities (obviously I don’t, to be honest it was really a good moment to build something like that).
We tested one or two models before dismantling the track.
On the December 19th, the track was reassembled on the final site. The test on the models did not go as well as previously; the room was not brightly lit enough. The competition looked complicated.
Two salespeople entertained and presented the show, all CBM discovered this small autonomous car that learns from its mistakes. There were the captivated people, IT connoisseurs and those who were more interested in the piña colada.
We started the race with the “VroomVroomRacing” team (Benjamin Houx, Rudy Sarrazin and Carlos Goncalves), all CBM staring at the car… which didn’t move for nearly two long, very long minutes, the brightness being not as good as it used to have. The first team that had never tested their model on a real circuit took a long time to complete a full lap, but they succeeded in doing so.
Next, the “TrainedForSpeed” team (Anthony Foulfoin, Nicolas Poulain and Christophe Colin), which had already tested its model on a real track did not get any better. Finally, the last team iRobotRacer(presented as the favorite with Fadwa Messaoudi, Florence Chabanois and Gabriel Plassard) which had already won an AWS award and which had preloaded 3 models and was hyper trained, did not even succeed to run a full lap!
Conclusion: Anything could happen! Nothing is won in advance.
So, the big winners are the “VroomVroomRacing” team members!
Here are a few words from our great winners:
“We wanted to base our model on how a real world driver would make decisions: he/she tries to compute the optimal car heading by calculating the average of next currently visible waypoints. Limiting it to waypoints visible by the camera reduces the risk of the car trying to cut through a turn.
The race itself was very different from our online practice as there are a lot of potential distractions for the car and we had to adjust its speed in real time. At first it had trouble finishing a single lap without going off limits, but once it started figuring out turns and walls we were able to get a very clean lap! It was very fun to put our algorithm to the test, especially since it went pretty well!”