What is Season Beta?
Even though we explained many of these points in our shows and on social media, there’s still a lot of evidence that this topic needs a systematic explanation. That’s why we came back here to Medium to give the answers in a single article.
Roborace is the world’s first driverless racing series. It’s an extreme competition of full-size racing robots where coding teams from the USA, UK, Austria, Italy, and Switzerland all go for the win.
In September 2020, Roborace started its test season titled “Season Beta” and introduced the real-time mixed reality feature named “Roborace Metaverse”. In every round of this season, the AI drivers developed by the competing teams race against virtual objects appearing on the actual race track. These objects have various forms and behaviour simply classified into obstacles that the AI driver should avoid and collectables that it should catch. The format of this competition is solo sprints in a time attack manner with a scoring system that affects result time depending on the number and severity of object hits. The hits of obstacles cause penalties increasing the result time. The hits or catches of collectables (aka loot) give bonuses reducing the result time. The rounds of the competition are arranged in missions with specific objectives and rules issued by Roborace for each mission in advance. Each mission introduces new kinds of objects, also increasing the dynamics and complexity of their behaviour.
In Mission 4 we also introduced the ghost cars that teams had to overtake for earning bonuses. Hitting a ghost car was scored as hitting an obstacle.
This is not a comprehensive variety of possible virtual objects. We are committed to introducing more and more of them in future missions. At a particular stage, we will also make them interactive, so the audience can have a chance to control them on the track during the races.
What is the reason for this format? What is the purpose of Metaverse?
Some people also ask why the Metaverse objects look so cartoonish. We’ll tell this in another article 😉
As many people expect from the races of self-driving machines, the cars should race against each other. And the more speedy they go and the more aggressively they interact with each other, the better and more spectacular the competition becomes. That’s what is called wheel-to-wheel racing. And in our strategy it’s the crucial capability for the teams to enter Season 1 in 2022.
High-speed overtaking is an essential element of wheel-to-wheel racing. But it’s not an easy task for self-driving AI and its developers. It implies a complex set of various things to be implemented before you can go full throttle. And the program of Season Beta is intentionally arranged in a way helping the competing teams to develop and validate all these elements to achieve sufficient performance expected to do real wheel-to-wheel racing.
What are the problems to be solved to get this achieved?
First, it’s sophisticated enough adaptive motion control. This thing aims to adjust the vehicle’s actual motion to the trajectory computed by the path planning algorithm of the self-driving system. Of course, you can pre-calculate a mathematically ideal trajectory and can even have certain chances to execute this trajectory without complex motion control on a perfect track. Still, the actual circumstances on a generic track are very stochastic. The traction is not ideal and may vary in different places of the track. The aero aspects play a lot at high speeds, especially when it’s windy. Plus, consider possible weather changes. So it would be best if you take that all into account when your code executes a trajectory of a full-size and relatively heavy vehicle (DevBot is 1300 kg) running at speeds above 100 kph. Our teams actually go up to 200 kph, and will go faster and faster in further missions.
Second, it’s dynamic path planning. This is about how often and efficiently you can re-calculate the possible trajectories and select the right one for responding to actual circumstances. When your car is alone on the track, you can even do the path rendering in advance and validate it in a simulator using relatively unlimited compute power. Doing this solely onboard in an autonomous car is another game, especially when talking about dynamic re-calculation responding to sudden circumstances. The computing power inside the vehicle is limited, and you share it with other critical tasks to perform for driving.
There are other aspects to take into account that may make this story even longer. For instance, you may also take care of the racing strategy overall, so your AI racer makes choices considering various possible consequences and their impact on the final result of the race. That all is what forms the performance of your extreme self-driving system and leads it to the win.
The usual question is whether this is relevant to the everyday applications of self-driving systems for public roads. Suppose you think about vehicle autonomy not as just a perk for a human driver to take some rest but rather as an opportunity to make the roads safer by giving cars superhuman capabilities. In that case, you can imagine how vital are the exercises that we do in Roborace when challenging self-driving systems to react to extreme conditions.
There’s an opinion that developing self-driving systems for public roads and race circuits are entirely different disciplines. That’s true. Public streets and highways are usually designed to be comfortable for drivers of any experience. There’re respective standards and regulations to enforce this. Race circuits are designed to be difficult. A self-driving system on a race track faces much more challenges than on a public road. Not just because of speed which is still an important aspect, but also because of way higher dynamics of possible scenarios. And we increase these dynamics even more by introducing Metaverse elements. So even when our teams achieve high-performance wheel-to-wheel racing capabilities, we will not switch off the Metaverse. Instead, we will raise more and more complex scenarios using Metaverse because challenging robots is a pleasure.
Why not just doing that all in a simulator? Of course, simulation is a great helper for developing self-driving systems. But even with cutting-edge tools, it’s still far from being 100% accurate. That’s why in Roborace the AI drivers run in mixed reality so that we can get the best of both worlds — real and virtual.
So what is the plan for Season Beta?
We arranged it in 3 stages having different levels of complexity. Further, I cite the official plan issued to the teams on June 23rd, 2020.
Here you can see mentioned the V2X system that informs cars about the objects and other vehicles. Currently, for this purpose we use the V2X system based on 802.11p radio and ETSI CAM messages.
Stage 1: Static Objects
1a: Zoned Objects
- Object zones are 300 m in length and are separated by clear track free zones, which are 300 m in length, resulting in three of each zone per lap of the testing track
- Objects are placed at random by Roborace in the next obstacle zone as the ego vehicle enters the current obstacle zone
- V2X sends all objects in current and next object zones
- Object placement in each zone may change each lap
- Objects can be either collectable or avoidable objects
- Objects are virtual
1b: Global Objects
- Zones are removed
- Objects can be placed randomly on track
- V2X sends all objects within an area of 100 m ahead, and 30 m behind the ego vehicle
- Object placement may change each lap
- Objects can be either collectable or avoidable objects
- Objects are virtual
Stage 2: Dynamic Objects
2a: Slow moving, fixed direction vector objects
- Objects may have non-zero velocity and acceleration
- Objects movement is limited to parallel to the track centreline
2b: Ghost Cars
- One or more competitive vehicles injected by Metaverse
- Driving line and relative performance of injecting vehicles can be incrementally more challenging (i.e. centreline only, LHS parallel to centreline, RHS parallel to centreline, racing line)
- Injected cars follow a fixed trajectory only (i.e. do not react to the ego vehicle)
Stage 3: Racing
3a: Mixed Reality
- Maximum one physical car on track versus one or more virtual cars in the simulator
- All cars driven in real-time by teams AI drivers
- All cars reacting to the other cars on track, whether physically or virtually
3b: Wheel-to-Wheel
- One or more physical cars on track versus none or more virtual cars in the simulator
- All cars driven in real-time by teams AI drivers
- All cars reacting to the other cars on track, whether physically or virtually
Even though we defined this plan months before Season Beta commenced, we didn’t design all its missions that far in advance. At the current stage, we finalise the next mission plan right after the preceding mission finishes to consider its results and learnings. The rationale for this is to get every next mission aligned with the actual achievements of our teams. This also gives us some freedom to improve the racing format from the findings we get while running these events and the feedback we get from the audience.
I hope this clarifies things somehow. With having this information while watching our live streams, you may judge where we are in Season Beta, what is achieved, and what needs further progress 🤓
Roborace team is eager to be very open with the community. If any questions and suggestions, please drop us a message at hello@roborace.com.
Watch Roborace live streams on Twitch: https://twitch.tv/roborace
Connect with Roborace online:
- YouTube: https://youtube.com/roborace
- Instagram: https://instagram.com/roborace
- Facebook: https://facebook.com/roborace
- Twitter: https://twitter.com/roborace
