Having spent the last years studying the art of AWS DeepRacer in the physical world, the author went to AWS re:Invent 2024. How did it go?
In AWS DeepRacer: How to master physical racing?, I wrote in detail about some aspects relevant to racing AWS DeepRacer in the physical world. We looked at the differences between the virtual and the physical world and how we could adapt the simulator and the training approach to overcome the differences. The previous post was left open-ended—with one last Championship Final left, it was too early to share all my secrets.
Now that AWS re:Invent is over, it’s time to share my strategy, how I prepared, and how it went in the end.
Strategy
Going into the 2024 season, I was reflecting on my performance from 2022 and 2023. In 2022, I had unstable models that were unable to do fast laps on the new re:Invent 2022 Championship track, not even making the last 32. In 2023, things went slightly better, but it was clear that there was potential to improve.
Specifically, I wanted a model that:
- Goes straight on the straights and corners with precision Has a survival instinct and avoids going off-track even in a tight spot Can ignore the visual noise seen around the track
Combine that with the ability to test the models before showing up at the Expo, and success seemed possible!
Implementation
In this section, I will explain my thinking about why physical racing is so different than virtual racing, as well as describe my approach to training a model that overcomes those differences.
How hard can it be to go straight?
If you have watched DeepRacer over the years, you have probably seen that most models struggle to go straight on the straights and end up oscillating left and right. The question has always been: why is it like that? This behavior causes two issues: the distance driven increases (result: slower lap time) and the car potentially enters the next turn in a way it can’t handle (result: off-track).
A few theories emerged:
- Sim-to-real issues – The steering response isn’t matching the simulator, both with regards to the steering geometry and latency (time from picture to servo command, as well as the time it takes the servo to actually actuate). Therefore, when the car tries to adjust the direction on the straight, it doesn’t get the response it expects. Model issues – A combination of the model not actually using the straight action, and not having access to angles needed to dampen oscillations (2.5–5.0 degrees). Calibration issues – If the car isn’t calibrated to go straight when given a 0-degree action, and the left/right max values are either too high (tendency to oversteer) or too low (tendency to understeer), you are likely to get control issues and unstable behavior.
My approach:
- Use the Ackermann steering geometry patch. With it, the car will behave more realistically, and the turning radius will decrease for a given angle. As a result, the action space can be limited to angles up to about 20 degrees. This roughly matches with the real car’s steering angle. Include stabilizing steering angles (2.5 and 5.0) in the action space, allowing for minor corrections on the straights. Use relatively slow speeds (0.8–1.3 m/s) to avoid slipping in the simulator. My theory is that the 15 fps simulator and the 30 fps car actually translates 1.2 mps in the simulator into effectively 2.4 mps in the real world. By having an inverted chevron action space giving higher speeds for straights, nudge the car to use the straight actions, rather than oscillating left-right actions. Try out v3, v4, and v5 physical models—test on a real track to see what works best. Otherwise, the reward function was the same progress-based reward function I also use in virtual racing.
The following figure illustrates the view of testing in the garage, going straight at least one frame.
Be flexible
Virtual racing is (almost) deterministic, and over time, the model will converge and the car will take a narrow path, reducing the variety in the situations it sees. Early in training, it will frequently be in odd positions, almost going off-track, and it remembers how to get out of these situations. As it converges, the frequency at which it must handle these reduces, and the theory is that the memory fades, and at some point, it forgets how to get out of a tight spot.
My approach:
- Diversify training to teach the car to handle a variety of corners, in both directions:
- Consistently train models going both clockwise and counterclockwise. Use tracks—primarily the 2022 Championship track—that are significantly more complex than the Forever Raceway. Do final optimization on the Forever Raceway—again in both directions.
Stay focused on the track
In my last post, I looked at the visual differences between the virtual and real worlds. The question is what to do about it. The goal is to trick the model into ignoring the noise and focus on what is important: the track.
My approach:
- Train in an environment with significantly more visual noise. The tracks in the custom track repository have added noise through additional lights, buildings, and different walls (and some even come with shadows). Alter the environment during training to avoid overfitting to the added noise. The custom tracks were made in such a way that different objects (buildings, walls, and lines) could be made invisible at runtime. I had a cron job randomizing the environment every 5 minutes.
The following figure illustrates the varied training environment.
What I didn’t consider this year was simulating blurring during training. I attempted this previously by averaging the current camera frame with the previous one before inferencing. It didn’t seem to help.
Lens distortion is a topic I have observed, but not fully investigated. The original camera has a distinct fish-eye distortion, and Gazebo would be able to replicate it, but it would require some work to actually determine the coefficients. Equally, I have never tried to replicate the rolling motions of the real car.
Testing
Testing took place in the garage on the Trapezoid Narrow track. The track is obviously basic, but with two straights and two 180-degree turns with different radii, it had to do the job. The garage track also had enough visual noise to see if the models were robust enough.
The method was straightforward: try all models both clockwise and counterclockwise. Using the logs captured by the custom car stack, I spent the evening looking through the video of each run to determine which model I liked the best—looking at stability, handling (straight on straights plus precision cornering), and speed.
re:Invent 2024
The track for re:Invent 2024 was the Forever Raceway. The shape of the track isn’t new; it shares the centerline with the 2022 Summit Speedway, but being only 76 cm wide (the original was 1.07 cm), the turns become more pronounced, making it a significantly more difficult track.
The environment
The environment is classic re:Invent: a smooth track with very little shine combined with smooth, fairly tall walls surrounding the track. The background is what often causes trouble—this year, a large lit display hung under the ceiling at the far end of the track, and as the following figure shows, it was attracting quite some attention from the GradCam.
Similarly, the pit crew cage, where cars are maintained, attracted attention.
The results
So where did I end up, and why? In Round 1, I ended up at place 14, with a best average of 10.072 seconds, and a best lap time of 9.335 seconds. Not great, but also not bad—almost 1 second outside top 8.
Using the overhead camera provided by AWS through the Twitch stream, it’s possible to create a graphical view showing the path the car took, as shown in the following figure.
If we compare this with how the same model liked to drive in training, we see a bit of a difference.
What becomes obvious quite quickly is that although I succeeded in going straight on the (upper) straight, the car didn’t corner as tightly as during training, making the bottom half of the track a bit of a mess. Nevertheless, the car demonstrated the desired survival instinct and stayed on track even when faced with unexpectedly sharp corners.
Why did this happen:
- 20 degrees of turning using Ackermann steering is too much; the real car isn’t capable of doing it in the real world The turning radius is increasing as the speed goes up due to slipping, caused both by low friction and lack of grip due to rolling The reaction time plays more of a role as the speed increases, and my model acted too late, overshooting into the corner
The combined turning radius and reaction time effect also caused issues at the start. If the car goes slowly, it turns much faster—and ends up going off-track on the inside—causing issues for me and others.
My takeaways:
- Overall, the training approach seemed to work well. Well-calibrated cars went straight on the straights, and background noise didn’t seem to bother my models much. I need to get closer to the car’s actual handling characteristics at speed during training by increasing the max speed and reducing the max angle in the action space. Physical racing is still not well understood—and it’s a lot about model-meets-car. Some models thrive on objectively perfectly calibrated cars, whereas others work great when matched with a particular one. Track is king—those that had access to the track, either through their employer or having built one at home, had a massive advantage, even if almost everyone said that they were surprised by which model worked in the end.
Now enjoy the inside view of a car at re:Invent, and see if you can detect any of the issues that I have discussed. The video was recorded after I had been knocked out of the competition using a car with the custom car software.
Closing time: Where do we go from here?
This section is best enjoyed with Semisonic’s Closing Time as a soundtrack.
As we all wrapped up at the Expo after an intense week of racing, re:Invent literally being dismantled around us, the question was: what comes next?
This was the last DeepRacer Championship, but the general sentiment was that whereas nobody will really miss virtual racing—it is a problem solved—physical racing is still a whole lot of fun, and the community is not yet ready to move on. Since re:Invent several initiatives have gained traction with a common goal to make DeepRacer more accessible:
- By enrolling cars with the DeepRacer Custom Car software stack into DeepRacer Event Manager you can capture car logs and generate the analytics videos, as shown in this article, directly during your event!
- DeepRacer Pi and DeepRacer Custom Car initiatives allow racers to build cars at home:Use off-the-shelf components for a 1:18 scale racer, or
DeepRacer Custom Console will be a drop-in replacement for the current car UI with a beautiful UI designed in Cloudscape, aligning the design with DREM and the AWS Console.
Prototype DeepRacer Pi Mini – 1 :28 scale
Closing Words
DeepRacer is a fantastic way to teach AI in a very physical and visual way, and is suitable for older kids, students, and adults in the corporate setting alike. It will be interesting to see how AWS, its corporate partners, and the community will continue the journey in the years ahead.
A big thank you goes to all of those that have been involved in DeepRacer from its inception to today—too many to be named—it has been a wonderful experience. A big congratulations goes out to this years’ winners!
Closing time, every new beginning comes from some other beginning’s end…
About the Author
Lars Lorentz Ludvigsen is a technology enthusiast who was introduced to AWS DeepRacer in late 2019 and was instantly hooked. Lars works as a Managing Director at Accenture, where he helps clients build the next generation of smart connected products. In addition to his role at Accenture, he’s an AWS Community Builder who focuses on developing and maintaining the AWS DeepRacer community’s software solutions.
