2026 Aston Martin Aramco Formula One Team

[ispano6]

There were lots of driver comments this year, explicitly stating that the AI / algorithm has been adjusting the deployment substantially based on where the driver lifted in the lap or corners before, so evidently it is constantly changing under changing scenarios.

As a person who works in training AI models and generative AI and ML, I can appreciate what teams may be doing with how they would be creating maps and use-case scenarios in how and when those maps are used. I'm fairly certain they aren't coming up with new maps during the race, that I know. ML will help identify during the practice sessions what the best maps to use will be for quali and race sims. Those will then be used during those sessions. Honda had been doing something similar already during the RedBull era with a dedicated PU team in a satellite location in Japan as they monitored telemetry and observed the PU output across the weekend and also in real time during the race. The 'algorithms' aren't rocket science but rather using the telemetry on the fly to assess how closely or far off they are from the idealized run plan. The goal and purpose of ML would not be to come up with the algorithm during the race but rather how closely they can operate the package with the expected variables to achieve convergence with their ML trained run plan.

[diffuser]

The scenario where at "300 km/h → ECU detects coasting → automatically selects neutral → charges battery" is not allowed by the rules. Neutral can only be selected by the driver. The ECU might be allowed to charge the battery in neutral above 50kph though.

Article C9.8.3 says:

Each individual gear change must be separately initiated by the driver.
A shift to gearbox neutral is specifically allowed as a multiple-change exception, but it must be “following a request from the driver.”
Automatic gear changes are classified as a driver aid and are not permitted.

C5.2.12 prohibits MGU-K use until 50 km/h

Also, you would hit the shifter to come out of neutral and that normally sticks you in 1st gear.

There's a lot of talk about algorithms but it seems like a misuse of the word or actual concept as it applied as if it is constantly changing while the driver is on the track and under changing conditions and scenarios. It could be considered a driver aid if the map is designed specifically for the circuit. So I'm thinking the driver has to choose a map that changes the operation but is otherwise consistent regardless of the track layout. I also think the FIA would have to heavily monitor the map usage and how the driver is instructed to use said map. It may be other teams that call it out before the FIA notices too.

There were lots of driver comments this year, explicitly stating that the AI / algorithm has been adjusting the deployment substantially based on where the driver lifted in the lap or corners before, so evidently it is constantly changing under changing scenarios.

The difference between an algorithm and a drivers aid, is the regs say one is a drivers aid the other is a algorithm...

[DDopey]

There were lots of driver comments this year, explicitly stating that the AI / algorithm has been adjusting the deployment substantially based on where the driver lifted in the lap or corners before, so evidently it is constantly changing under changing scenarios.

As a person who works in training AI models and generative AI and ML, I can appreciate what teams may be doing with how they would be creating maps and use-case scenarios in how and when those maps are used. I'm fairly certain they aren't coming up with new maps during the race, that I know. ML will help identify during the practice sessions what the best maps to use will be for quali and race sims. Those will then be used during those sessions. Honda had been doing something similar already during the RedBull era with a dedicated PU team in a satellite location in Japan as they monitored telemetry and observed the PU output across the weekend and also in real time during the race. The 'algorithms' aren't rocket science but rather using the telemetry on the fly to assess how closely or far off they are from the idealized run plan. The goal and purpose of ML would not be to come up with the algorithm during the race but rather how closely they can operate the package with the expected variables to achieve convergence with their ML trained run plan.

It does look like some live parameters that are given to a ml algorithm can heavily influence the deployment during a lap. It doesn’t have to be learning during race, but outcomes still depends on which parameters you use. Based on the responses by drivers it doesn’t look like fixed precaculated maps.

[collindsilva]

How much laptime is the Honda upgrade rumored to bring? I'd guess the bump in regen to the max should be sizeable in itself.

It depends on how much deficit Honda has reduced on the ICE alone, if they have managed to reduce the deficit by more than 50% say about 30HP, that should give them at least .5 seconds (based on various rumours on X), now add improved regeneration, chassis optimization, gearbox optimization etc, should be minimum 1 sec.

[aMessageToCharlie]

There were lots of driver comments this year, explicitly stating that the AI / algorithm has been adjusting the deployment substantially based on where the driver lifted in the lap or corners before, so evidently it is constantly changing under changing scenarios.

As a person who works in training AI models and generative AI and ML, I can appreciate what teams may be doing with how they would be creating maps and use-case scenarios in how and when those maps are used. I'm fairly certain they aren't coming up with new maps during the race, that I know. ML will help identify during the practice sessions what the best maps to use will be for quali and race sims. Those will then be used during those sessions. Honda had been doing something similar already during the RedBull era with a dedicated PU team in a satellite location in Japan as they monitored telemetry and observed the PU output across the weekend and also in real time during the race. The 'algorithms' aren't rocket science but rather using the telemetry on the fly to assess how closely or far off they are from the idealized run plan. The goal and purpose of ML would not be to come up with the algorithm during the race but rather how closely they can operate the package with the expected variables to achieve convergence with their ML trained run plan.

While you're probably right and I do agree that details are important - this is mostly semantics at this point. Whatever you might call these reactive, on the fly software adjustments, they shoud clearly be classified as a driver aid IMO.

But none of this is specific to AM, so let's keep it at that.

[etusch]

I hate this cause internal targets could mean anything from 60hp to "it runs"

It's about looking at things positively and putting the pieces together.
Positive: we achieved our goals,
goals like faster combustion and durability, a big improvement rather than small steps. Yes, there's some uncertainty, but not for Aston Honda, but for us and the other teams (for now). For them, it's not uncertain.
Faster combustion and durability might not be so uncertain. You also commented on the big step: Ice; 30 hp, plus another 30 from better electric regeneration. We can consider this as big step. If they meant 25 hp or 35 hp by saying big step, only time will tell.

[selvam_e2002]

How much lag we have with other engines hp? I don't think it is not >100 HP, if they even improve 35 to 50hp in season, it is very huge. I never come across any F1 team improved 2.5 seconds in one upgrade. So I think AMR in right direction. If they can fight for podium end of this year, then the next year is WDC and WCC year for AMR.

[selvam_e2002]

Also who knows, they have rocket engine to go for Podium in Netherlands....

[diffuser]

There were lots of driver comments this year, explicitly stating that the AI / algorithm has been adjusting the deployment substantially based on where the driver lifted in the lap or corners before, so evidently it is constantly changing under changing scenarios.

As a person who works in training AI models and generative AI and ML, I can appreciate what teams may be doing with how they would be creating maps and use-case scenarios in how and when those maps are used. I'm fairly certain they aren't coming up with new maps during the race, that I know. ML will help identify during the practice sessions what the best maps to use will be for quali and race sims. Those will then be used during those sessions. Honda had been doing something similar already during the RedBull era with a dedicated PU team in a satellite location in Japan as they monitored telemetry and observed the PU output across the weekend and also in real time during the race. The 'algorithms' aren't rocket science but rather using the telemetry on the fly to assess how closely or far off they are from the idealized run plan. The goal and purpose of ML would not be to come up with the algorithm during the race but rather how closely they can operate the package with the expected variables to achieve convergence with their ML trained run plan.

It does look like some live parameters that are given to a ml algorithm can heavily influence the deployment during a lap. It doesn’t have to be learning during race, but outcomes still depends on which parameters you use. Based on the responses by drivers it doesn’t look like fixed precaculated maps.

That is not how I understand it. The driver selects a MAP, which essentially defines a preferred energy-management strategy. Since the PUs cannot regenerate enough energy to allow maximum deployment whenever the driver wants it, compromises have to be made regarding when and where to deploy energy, when to harvest/recharge, and by how much.

The difference between one MAP and another is essentially where on the lap they prioritise deployment and harvesting, and by how much. In that sense, the MAPs provide a set of targets or instructions for the energy-management algorithms.

The algorithms then manage the actual deployment and harvesting within the constraints of the regulations—i.e. what is permitted—according to the demands of the selected MAP and the actual throttle input. So, the MAP is not directly controlling every deployment or recharge event; rather, it defines the strategy that the algorithms try to achieve.

In a race, for example, if you want to pass someone, you might notice that the car ahead is weaker on the exit of Turn 4. There might be a MAP that prioritises deployment on the exit of that turn, giving you more energy to use where you need it to make the pass.

[DDopey]

As a person who works in training AI models and generative AI and ML, I can appreciate what teams may be doing with how they would be creating maps and use-case scenarios in how and when those maps are used. I'm fairly certain they aren't coming up with new maps during the race, that I know. ML will help identify during the practice sessions what the best maps to use will be for quali and race sims. Those will then be used during those sessions. Honda had been doing something similar already during the RedBull era with a dedicated PU team in a satellite location in Japan as they monitored telemetry and observed the PU output across the weekend and also in real time during the race. The 'algorithms' aren't rocket science but rather using the telemetry on the fly to assess how closely or far off they are from the idealized run plan. The goal and purpose of ML would not be to come up with the algorithm during the race but rather how closely they can operate the package with the expected variables to achieve convergence with their ML trained run plan.

It does look like some live parameters that are given to a ml algorithm can heavily influence the deployment during a lap. It doesn’t have to be learning during race, but outcomes still depends on which parameters you use. Based on the responses by drivers it doesn’t look like fixed precaculated maps.

That is not how I understand it. The driver selects a MAP, which essentially defines a preferred energy-management strategy. Since the PUs cannot regenerate enough energy to allow maximum deployment whenever the driver wants it, compromises have to be made regarding when and where to deploy energy, when to harvest/recharge, and by how much.

The difference between one MAP and another is essentially where on the lap they prioritise deployment and harvesting, and by how much. In that sense, the MAPs provide a set of targets or instructions for the energy-management algorithms.

The algorithms then manage the actual deployment and harvesting within the constraints of the regulations—i.e. what is permitted—according to the demands of the selected MAP and the actual throttle input. So, the MAP is not directly controlling every deployment or recharge event; rather, it defines the strategy that the algorithms try to achieve.

In a race, for example, if you want to pass someone, you might notice that the car ahead is weaker on the exit of Turn 4. There might be a MAP that prioritises deployment on the exit of that turn, giving you more energy to use where you need it to make the pass.

Bit difficult with so many unknowns to talk about this. But for me this would not explain why drivers at the same point in the track get a different deployment than they expect. So for me this would be explained by some type of algorithm that would have gotten different parameters at that same point and decided to distribute the power differently.

[diffuser]

It does look like some live parameters that are given to a ml algorithm can heavily influence the deployment during a lap. It doesn’t have to be learning during race, but outcomes still depends on which parameters you use. Based on the responses by drivers it doesn’t look like fixed precaculated maps.

That is not how I understand it. The driver selects a MAP, which essentially defines a preferred energy-management strategy. Since the PUs cannot regenerate enough energy to allow maximum deployment whenever the driver wants it, compromises have to be made regarding when and where to deploy energy, when to harvest/recharge, and by how much.

The difference between one MAP and another is essentially where on the lap they prioritise deployment and harvesting, and by how much. In that sense, the MAPs provide a set of targets or instructions for the energy-management algorithms.

The algorithms then manage the actual deployment and harvesting within the constraints of the regulations—i.e. what is permitted—according to the demands of the selected MAP and the actual throttle input. So, the MAP is not directly controlling every deployment or recharge event; rather, it defines the strategy that the algorithms try to achieve.

In a race, for example, if you want to pass someone, you might notice that the car ahead is weaker on the exit of Turn 4. There might be a MAP that prioritises deployment on the exit of that turn, giving you more energy to use where you need it to make the pass.

Bit difficult with so many unknowns to talk about this. But for me this would not explain why drivers at the same point in the track get a different deployment than they expect. So for me this would be explained by some type of algorithm that would have gotten different parameters at that same point and decided to distribute the power differently.

It wasn't meant to explain why drivers get different deployments than they expect. It was just my understanding of what the algorithms control and what they don't control.

I'm not sure this is true for every single deployment anomaly we've seen or heard about, but my understanding is that the issue is that a small change in what the driver did led to a deployment that the driver didn't intend or want. It wasn't that tge driver did the EXACT same thing.

As a programmer, I can say that's just how programs react when you haven't defined what should happen in a very specific situation. They tend to do something unexpected. So I chalk that up immature code. As the code gets better it gets more experienced, there will be fewer exceptions.

That is MHO.

[ispano6]

While you're probably right and I do agree that details are important - this is mostly semantics at this point. Whatever you might call these reactive, on the fly software adjustments, they shoud clearly be classified as a driver aid IMO.

But none of this is specific to AM, so let's keep it at that.

I for one think ML during a race IS a driver AID and should be banned. Everything I said points to how it should be used, not how they are using it. You are not allowed to upload a new "algorithm" to the car while it is running and you are not allowed to let the car automatically adjust deployment during a lap.
That is blatant driver aid.

[ispano6]

It does look like some live parameters that are given to a ml algorithm can heavily influence the deployment during a lap. It doesn’t have to be learning during race, but outcomes still depends on which parameters you use. Based on the responses by drivers it doesn’t look like fixed precaculated maps.

Again, the algorithm determines when to use what map, it doesn't create new maps on the fly for the driver to use. It doesn't change maps for the driver either. The algorithm tells trackside personnel whether or not to deviate from the pre-designated run plan, which consists of the usage of different maps for changing parameters and variables during the race. The algorithm can't create a new map and upload it to the car. It can't automatically change maps during deployment without driver input. The drivers don't like when the algorithm tells the trackside personnel, that then relay which mode to use, when it seems counter-intuitive to the driver to use a map that deviates from the established runplan. I think you're mixing up maps and runplan. you can change the runplan, you can't create new maps.

[diffuser]

It does look like some live parameters that are given to a ml algorithm can heavily influence the deployment during a lap. It doesn’t have to be learning during race, but outcomes still depends on which parameters you use. Based on the responses by drivers it doesn’t look like fixed precaculated maps.

Again, the algorithm determines when to use what map, it doesn't create new maps on the fly for the driver to use. It doesn't change maps for the driver either. The algorithm tells trackside personnel whether or not to deviate from the pre-designated run plan, which consists of the usage of different maps for changing parameters and variables during the race. The algorithm can't create a new map and upload it to the car. It can't automatically change maps during deployment without driver input. The drivers don't like when the algorithm tells the trackside personnel, that then relay which mode to use, when it seems counter-intuitive to the driver to use a map that deviates from the established runplan. I think you're mixing up maps and runplan. you can change the runplan, you can't create new maps.

The driver chooses MAP. During a race Alonso and Stroll predominant use MAPs 4, 6 and 7. That doesn't mean much cause the MAP aren't the same from track to track. MAPs are track specific. They use another MAP for Qualifying.

[GoranF1]

Reports of problems whit the PU during filming day.