AWS DeepRacer League is a global race where programmable DeepRacers, i.e. autonomous vehicles created by Amazon Web Services (AWS), compete for the title of the fastest. The winner is the first one to learn to overcome obstacles faced along the route. Can you get to Las Vegas by this vehicle? It turns out you can.
What’s the harm in trying?
When we applied for the AWS DeepRacer League, we thought it would be great if we made it to the top fifty. After all, competitors from all over the world take part in the race, says Marcin Zima, Java Team Leader from the Rzeszów PGS Software office.
Tomek Panek, a Java Developer from Rzeszów, was the guy who dedicated his time and attention to develop DeepRacer. It quickly turned out that he was a perfect match for this job. His previous experience in the Machine Learning area helped, and the DeepRacer project was to be the next step in his career.
– It was hard at the beginning because it was the first such a league in history so it was new to everyone -says Tomek Panek. We joined the competition in May 2019. For the first 2-3 weeks of the race, we didn’t hope for too much, but we broke it down.
The race was held in a virtual environment that was to map the actual track. The AWS ready-made service could be used for machine learning. However, in order to compete successfully, it was necessary to adopt extraordinary measures and build the entire infrastructure locally.
– It was a challenge – says Tomek reminiscently. I had to get to grips with all of it and learn more about DeepRacer, Reinforcement Learning or Amazon Web Services. AWS offered a tool that helped me set the speed and steering range, but its use was limited. There was no choice but to tackle the problem with a range of measures to freely modify these parameters. I had to learn it from scratch.
The first routes were not too complicated, but one by one they were getting more and more difficult. The first models that Tomek developed operated at a constant speed due to small number of bends. However, each subsequent route required a vehicle that slows down before the bend and predicts what will happen on the route.
Finally, the team was a success. In August, Tomek Panek came first in the virtual league.
I was relieved that something cool happened and I shyly hoped I could qualify for the finals – he recalls. I was also stressed out a bit because I had to hold on to the lead and the differences between the first eighteen competitors, who secured themselves a trip to the Las Vegas finals, were minimal.
In fact, it is decimal places we are talking about here. Such tiny differences could be caused by many factors, e.g. the vehicle might have turned a little on a straight line at some point and precious fractions of a second were gone. However, the more complicated the route, the bigger the difference. The final eighteen seems to prove it: the first four competitors were neck and neck, and the final ones were several dozen seconds behind.
The virtual league winner could square off against the best DeepRacer competitors live in Las Vegas at the AWS re:Invent conference in December 2019. Those who get through to the finals included participants from all over the world, selected from the virtual league, and individuals who had the fastest time during races held at the AWS Summit conferences. In a nutshell: the real DeepRacer masters.
Tomek qualified for the Las Vegas finals from the second place in the ranking. The virtual league competition was extremely fierce. The difference between the first and the second place was 0.41 point. The race was divided into 6 stages with 1,000 points to get at each stage, less the lap time. Tomek scored 5,937.82 points, leaving hundreds of players from around the world behind him.
Viva Las Vegas
Tomek’s strategy for the final races was based on two driving modes: the first one was to drive very fast on the AWS reference route, and the other was to drive decently but on many different routes.
AWS re:Invent was a big shock. I didn’t really know what to expect. Some competitors practiced on their own track that mapped the final route. I was practising only virtually. My first thought was: it would be cool if the car did at least one solid lap – says Tomek, laughing. The difference between virtual reality and real-life racing was dramatic. I had five models: four of them to race on one route and an emergency model, which was a bit slower on the virtual track. In the morning on the competition day I tested them, and it turned out that all four models came out poorly and fall off the track. I was left with a generic model. It did fairly well in the virtual environment, but I didn’t expect it to produce good results. To my surprise it handled both the track and the bends skilfully. The competition organizers provided display units that showed how effectively the cars would take the bends and my car was in the lead throughout the entire competition. It just kept going and did not fall off the track. All I had to do was to set the right speed of the car so that it had good time. I had to experiment a little.
64 people divided into 4 groups of 16 competitors qualified for the first stage of the competition. The first four individuals from each group made it to the next stage, the rest dropped out. Each participant had 4 tracks and 4 minutes per each of them, a total of 16 minutes. The best time was considered the final score. Tomek qualified for the second stage from the fourth place.
Those who got through to the next stage knew exactly what they were doing. Most of them had attended the AWS Summits before or had their own tracks built. The competitors were neck and neck with about 0.5 second separating them. The level was higher and it was difficult to see what could be improved in the next lap, because such differences are unnoticeable. I was hoping luck would be on my side as a moment of weakness on the track puts an end to the race. Unfortunately, I was eliminated at this stage but I’m glad that I was ranked closer to the centre than the end of the ranking. – says Tomek. It was quite an adventure. I am immensely pleased with my result, because what happened in Vegas was totally beyond my expectations. What is more, I met a lot of knowledgeable people. After all, Reinforcement Learning is not a hot topic yet. We all learn it practically from scratch but the DeepRacer experience and the contacts we now have around the world, give us a good start.