T0@st
News Editor
- Joined
- Mar 7, 2023
- Messages
- 2,979 (3.82/day)
- Location
- South East, UK
| System Name | The TPU Typewriter |
|---|---|
| Processor | AMD Ryzen 5 5600 (non-X) |
| Motherboard | GIGABYTE B550M DS3H Micro ATX |
| Cooling | DeepCool AS500 |
| Memory | Kingston Fury Renegade RGB 32 GB (2 x 16 GB) DDR4-3600 CL16 |
| Video Card(s) | PowerColor Radeon RX 7800 XT 16 GB Hellhound OC |
| Storage | Samsung 980 Pro 1 TB M.2-2280 PCIe 4.0 X4 NVME SSD |
| Display(s) | Lenovo Legion Y27q-20 27" QHD IPS monitor |
| Case | GameMax Spark M-ATX (re-badged Jonsbo D30) |
| Audio Device(s) | FiiO K7 Desktop DAC/Amp + Philips Fidelio X3 headphones, or ARTTI T10 Planar IEMs |
| Power Supply | ADATA XPG CORE Reactor 650 W 80+ Gold ATX |
| Mouse | Roccat Kone Pro Air |
| Keyboard | Cooler Master MasterKeys Pro L |
| Software | Windows 10 64-bit Home Edition |
Monster Energy Supercross 25 - The Official Video Game—available as Dirt Master Edition or and Special Edition—is out now in early access for Xbox Series X|S, and the Standard Edition will be available for everyone on April 10, 2025—on Steam, Epic Games Store, and PlayStation 5. I can't wait to see all of you take the route for the Supercross elite in this new and completely improved game. Today, I want to talk to you about one of the most ambitious and challenging aspects of the game, which made us sweat but filled us with pride once the results started to arrive: Neural AI.
We started using our Neural AI, A.N.N.A (Artificial Neural Network Agent), in our games starting from 2019, but research began before then. Neural AI is a specific kind of AI that can simulate realistic behavior. To do that, it must train and test the tracks to learn the best trajectories to become as fast as possible. Over the years in Milestone we have seen incredible results, with AI capable of controlling vehicles in a realistic manner. But up to this Supercross chapter, A.N.N.A. has only been on tarmac surfaces.
Teaching an AI to Run in a New Environment
As mentioned before, A.N.N.A. was used in the past in the MotoGP and RIDE franchises, to race in situations that were intense and crowded. But tracks were made of tarmac, a very regular surface. Supercross tracks are much more compressed, short, and full of tight corners. They have a completely different rhythm thanks to the variety of jumps, like steps up, tabletops, walls, and more. Each jump has its own best way to do it, and once they are inserted on a track, riders must figure out the order of things and adjust based on what comes after those jumps, with the goal of being the fastest.
When we had to set the AI to handle all the different elements that make up Supercross tracks, we were worried it wouldn't be able to deal with all the variables it had to calculate during a lap around a dirt track. But right after the first tests, we knew that we were in the right direction. Programming a Neural AI takes a lot of time, and implementing A.N.N.A. in Supercross 25 is one of the factors that made us take two years to develop the next game.
How We Made It
When our AI is running during a race, it has to take into account a bunch of different factors, like track design, speed, balance, trajectory, other AI controlled riders, and, most importantly, players, who are the main unpredictable element on the track. It can handle all that complexity thanks to a lot of rigorous training. We wanted to include in the game the current Supercross season for the first time, but we had to start working on the AI way earlier due to timing reasons.
First, we calibrated the Agents on the new physics system, then we introduced tracks from the past to the AI so it could learn what a Supercross track is and what makes it up. To achieve the best level of realism, we immediately added an extra obstacle to our Agents: dynamic terrain. In real Supercross, ruts are deep furrows in the ground created by the bikes' passage. They change the feel and the looks of the track significantly. Usually, if you change one thing about a track after the AI has learned it, it can't find the original layout, and it gets lost. But with dynamic ruts, the tracks evolve every lap, so we had to teach A.N.N.A. how to handle this extra challenge, and it responded really well. Now it can adjust its trajectory and strategy based on the ruts present at that specific moment of the race.
When the 2025 Supercross tracks arrived, we made A.N.N.A. run on them, and we had amazing results. It was already a Supercross expert, able to handle and interpret the modules and variations of the tracks. The only exception was with elements it had never seen before, because every year the sport evolves, and track designers introduce original and interesting sequences. We had to teach it those extra modules, even making it go against its previous learnings. We were so impressed by the results we have achieved with this training system. Of course, we still made some fine-tuning, especially adjusting the difficulty of the AI based on the player's skill, but we had confirmation of how intelligent A.N.N.A. is, and this Supercross chapter is just the beginning for even more spectacular achievements.
See A.N.N.A. Flying
The air section is a huge part of Supercross, and in Supercross 25, we gave to both players and AI full control of the in-air motion thanks to the dual stick, which allows managing both the rider's and the bike's weight. We have taught A.N.N.A. the technique and made it train on all kinds of jumps. We were surprised when we saw that it could find and apply the best technique on a jump before we could! For example, there was a jump where the quickest way to make it was through a scrub—a particular bike maneuver that slices the air and allows shorter airtime. While I was still trying to understand how to perfect this trick for maximum efficiency, A.N.N.A. had already figured it out, with a top-notch execution. This has pushed us to become even better while playing against it!
Test It With Your Tracks
Neural AI is going to be available also on UGC tracks for the first time, and we have applied the same principle we used for the 2025 season. The Agents are smart and, as I like to say, they have studied a lot of textbooks. They have the knowledge and the ability to adapt to a new environment using the skills they have obtained through training just like a real rider. We are super happy with how they have responded, and we will keep working to make them even more prepared and precise.
Thank you for your time. It's always great to talk about these amazing achievements. We can't wait to see what spectacular stunts you will pull off. See you on the track!
Sebastiano Stefanetto, Game Director, Milestone
View at TechPowerUp Main Site | Source
We started using our Neural AI, A.N.N.A (Artificial Neural Network Agent), in our games starting from 2019, but research began before then. Neural AI is a specific kind of AI that can simulate realistic behavior. To do that, it must train and test the tracks to learn the best trajectories to become as fast as possible. Over the years in Milestone we have seen incredible results, with AI capable of controlling vehicles in a realistic manner. But up to this Supercross chapter, A.N.N.A. has only been on tarmac surfaces.
Teaching an AI to Run in a New Environment
As mentioned before, A.N.N.A. was used in the past in the MotoGP and RIDE franchises, to race in situations that were intense and crowded. But tracks were made of tarmac, a very regular surface. Supercross tracks are much more compressed, short, and full of tight corners. They have a completely different rhythm thanks to the variety of jumps, like steps up, tabletops, walls, and more. Each jump has its own best way to do it, and once they are inserted on a track, riders must figure out the order of things and adjust based on what comes after those jumps, with the goal of being the fastest.
When we had to set the AI to handle all the different elements that make up Supercross tracks, we were worried it wouldn't be able to deal with all the variables it had to calculate during a lap around a dirt track. But right after the first tests, we knew that we were in the right direction. Programming a Neural AI takes a lot of time, and implementing A.N.N.A. in Supercross 25 is one of the factors that made us take two years to develop the next game.
How We Made It
When our AI is running during a race, it has to take into account a bunch of different factors, like track design, speed, balance, trajectory, other AI controlled riders, and, most importantly, players, who are the main unpredictable element on the track. It can handle all that complexity thanks to a lot of rigorous training. We wanted to include in the game the current Supercross season for the first time, but we had to start working on the AI way earlier due to timing reasons.
First, we calibrated the Agents on the new physics system, then we introduced tracks from the past to the AI so it could learn what a Supercross track is and what makes it up. To achieve the best level of realism, we immediately added an extra obstacle to our Agents: dynamic terrain. In real Supercross, ruts are deep furrows in the ground created by the bikes' passage. They change the feel and the looks of the track significantly. Usually, if you change one thing about a track after the AI has learned it, it can't find the original layout, and it gets lost. But with dynamic ruts, the tracks evolve every lap, so we had to teach A.N.N.A. how to handle this extra challenge, and it responded really well. Now it can adjust its trajectory and strategy based on the ruts present at that specific moment of the race.
When the 2025 Supercross tracks arrived, we made A.N.N.A. run on them, and we had amazing results. It was already a Supercross expert, able to handle and interpret the modules and variations of the tracks. The only exception was with elements it had never seen before, because every year the sport evolves, and track designers introduce original and interesting sequences. We had to teach it those extra modules, even making it go against its previous learnings. We were so impressed by the results we have achieved with this training system. Of course, we still made some fine-tuning, especially adjusting the difficulty of the AI based on the player's skill, but we had confirmation of how intelligent A.N.N.A. is, and this Supercross chapter is just the beginning for even more spectacular achievements.
See A.N.N.A. Flying
The air section is a huge part of Supercross, and in Supercross 25, we gave to both players and AI full control of the in-air motion thanks to the dual stick, which allows managing both the rider's and the bike's weight. We have taught A.N.N.A. the technique and made it train on all kinds of jumps. We were surprised when we saw that it could find and apply the best technique on a jump before we could! For example, there was a jump where the quickest way to make it was through a scrub—a particular bike maneuver that slices the air and allows shorter airtime. While I was still trying to understand how to perfect this trick for maximum efficiency, A.N.N.A. had already figured it out, with a top-notch execution. This has pushed us to become even better while playing against it!
Test It With Your Tracks
Neural AI is going to be available also on UGC tracks for the first time, and we have applied the same principle we used for the 2025 season. The Agents are smart and, as I like to say, they have studied a lot of textbooks. They have the knowledge and the ability to adapt to a new environment using the skills they have obtained through training just like a real rider. We are super happy with how they have responded, and we will keep working to make them even more prepared and precise.
Thank you for your time. It's always great to talk about these amazing achievements. We can't wait to see what spectacular stunts you will pull off. See you on the track!
Sebastiano Stefanetto, Game Director, Milestone
View at TechPowerUp Main Site | Source
