2026 Hybrid Powerunits

[hollus]

I am pretty sure this is not AI in the sense of what passes for AI in the last 3 years.
“The algorithm” here is probably not more than a meduim complexity excel sheet, only running in real time and not on Excel.

[DDopey]

Before this generative AI that is now the hype, there was already AI, it's an umbrella term. Under it you have machine learning.
In ML you have models like, machine learning regression models. These take in a lot of parameters and then do a prediction of a value. This does not cost a lot of computation and is deterministic so this could be used. This makes most sense for the engines.

There are also a bit more of advanced ML models with deep learning, like Recurrent Neural Network (RNN) or Convolution Neural Network (CNN), depending on the complexity of the models these can also be used but require much more computation (still possible on embedded hardware, but heavier than the regression models). These are used to find hidden patterns, do classifications, image recognition and such. So depending on how sophisticated the need is this could also be used, but I dont think so.

[liuzka]

I was just wondering that everyone is complaining about AI and how they do not understand it, but that is bit stupid if it’s their own work.

While ML needs usually quite big dataset I would imagine that they run it in simulator before weekend as for base “map” and then it gets fine tuned during actual driving. If it works like this it might make simulator work very important and also that simulator driving is as close to real driving would be crucial. I have no idea if simulator is driven by computer or human.


On the other hand if customer teams get their “AI” from PU manufacturer then it is totally valid criticism.

[spokes]

Just putting my boring solution to all these "algorithm" "ai" deployment issues into the ether...here's to hoping someone at the FIA makes this the rule

I think this should, and could, be solved by having power modes that are closely related to the driver's throttle pedal position, and limit the amount of modes allowed (let's say 10). This way, engineers can work on developing deployment modes specific to tracks, but the driver controls which mode he uses.

For example, you have power modes 1-10, mode 1 deploys the least amount of energy, but inversely, harvests the most. therefore focuses on harvesting. Mode "5" delivers a balanced amount of energy deployment, and recharges an equal amount, allowing drivers to charge up what the same amount they use in a lap. Mode 10 would provide the most power, but not to exceed what the driver's right foot requests.

If the driver is giving 100% throttle he's getting 100% battery deployment, or if the driver is giving 50% throttle he's giving 50% of the battery deployment etc.

In this way, engineers can setup the modes specific to the track conditions, and drivers can get dependable output, allowing them to change modes when they need something from the battery/power unit.

[venkyhere]

Just putting my boring solution to all these "algorithm" "ai" deployment issues into the ether...here's to hoping someone at the FIA makes this the rule

I think this should, and could, be solved by having power modes that are closely related to the driver's throttle pedal position, and limit the amount of modes allowed (let's say 10). This way, engineers can work on developing deployment modes specific to tracks, but the driver controls which mode he uses.

For example, you have power modes 1-10, mode 1 deploys the least amount of energy, but inversely, harvests the most. therefore focuses on harvesting. Mode "5" delivers a balanced amount of energy deployment, and recharges an equal amount, allowing drivers to charge up what the same amount they use in a lap. Mode 10 would provide the most power, but not to exceed what the driver's right foot requests.

If the driver is giving 100% throttle he's getting 100% battery deployment, or if the driver is giving 50% throttle he's giving 50% of the battery deployment etc.

In this way, engineers can setup the modes specific to the track conditions, and drivers can get dependable output, allowing them to change modes when they need something from the battery/power unit.

My friend, the problem in question is not 'the quantum' (rate/duration) of charge/discharge , but the 'timing' of it => 'how-much' is not the issue, 'when' is the issue. The throttle pedal is a 'suggestion' now, rather than the 'command' it used to be before. The timing is decided by a computer based on how it 'senses' the position of the car on track (GPS ? distance covered ? custom positioning signal ? I'm not sure, but whatever it is, it's based on some 'learning' from previously fed data, in tandem with actual data collected during practice).

Its the most f****ed-up way to implement ANY RACING series.

[spokes]

My friend, the problem in question is not 'the quantum' (rate/duration) of charge/discharge , but the 'timing' of it => 'how-much' is not the issue, 'when' is the issue. The throttle pedal is a 'suggestion' now, rather than the 'command' it used to be before. The timing is decided by a computer based on how it 'senses' the position of the car on track (GPS ? distance covered ? custom positioning signal ? I'm not sure, but whatever it is, it's based on some 'learning' from previously fed data, in tandem with actual data collected during practice).

Its the most f****ed-up way to implement ANY RACING series.

I love the way you put it! a suggestion versus a command.
Because of the nature of electric power (being able to provide instant torque), my suggestion was more so a knob on the driver's steering wheel that essentially controls how much output can be given, relative to the driver's pedal.

i.e. in Mode 1, you only ever deploy max 10% of battery output at a given time. So if you give 100% throttle, in mode 1, you only get 10% of the batteries max deployment. If you do 50% throttle (still in mode one), you only deploy 5% of the batteries max deployment. The driver still commands how much power he wants, but he has more control over how much torque can be given depending on the mode he is in.

[ispano6]

I was just wondering that everyone is complaining about AI and how they do not understand it, but that is bit stupid if it’s their own work.

While ML needs usually quite big dataset I would imagine that they run it in simulator before weekend as for base “map” and then it gets fine tuned during actual driving. If it works like this it might make simulator work very important and also that simulator driving is as close to real driving would be crucial. I have no idea if simulator is driven by computer or human.


On the other hand if customer teams get their “AI” from PU manufacturer then it is totally valid criticism.

Correct, most of people who talkdown "AI" don't know what LLM, generative AI, ML really mean.
The core algorithm is a machine learning model. I've had to explain this in other places but most people here haven't taken a class on machine learning or have trained their own models so they don't appreciate what AI is doing.

ML models rely on polynomial equations that iterate over many data points to plot a line of best fit and given the number of factors and how much weight is applied to them (as exponents) becomes far more complex than a simple line of best fit. Over many iterations the algorithm/graph's inflection points reach convergence after having "mapped" a finite but large set of possibilities based on or accounting for factors that occur during the race - both predictable and the intangibles. The strategy is choose the model(s) or in this case "maps" that are best suited given the changing conditions.

[gearboxtrouble]

ML models aren't technically AI in my book. We've been using them for 20+ years in the financial industry and they're really just more dynamic statistical models - think regressions with benefits. Using them to optimize energy in an energy constrained formula is a no brainer and they were almost certainly used in the last engine rules as well. The difference is sub optimal behavior is far more visible when you're talking about a loss of 450hp instead of a loss of 80hp in the old rules (net the 80hp the MGUH could send to the K at full throttle). This feels like another red herring to divert attention away from the fundamental problem of this ruleset - the ICE component is undersized relative to the electrical component. No amount of software fiddling will change that and even the painful, long drawn out move to 60/40 will only help on the margins at places like Spa and Silverstone. To at least recover some driver agency they could maybe replace automated deployment with a two stage accelerator pedal where there is an easy and repeatable way to control when energy is deployed for the driver. We'll still be this energy starved disaster but at least when something goes awry it will be on the driver and not some computer.

[AR3-GP]

To at least recover some driver agency they could maybe replace automated deployment with a two stage accelerator pedal where there is an easy and repeatable way to control when energy is deployed for the driver. We'll still be this energy starved disaster but at least when something goes awry it will be on the driver and not some computer.

The current algorithms (although imperfect) have a side effect of homogenizing deployment across all the cars (eventually) to get optimal laptime. This has a hidden safety benefit. It reduces the frequency of big speed differences.

So ironically allowing a "free for all" electrical deployment by the driver is going to increase the frequency of large speed differences, and inevitably more accidents like the one of Bearman in Japan. Drivers simply would never be able to anticipate the butterfly effect of 1 extra second of deployment in the wrong place.

As you alluded to, its just the basic concept of a cars which can drive with either 500hp or 1000hp on the same pedal position due to the overzealous power split that is causing the current problems.


It should have never become "controversial" for the throttle pedal to have a single power output level that is scaled only by the position of the drivers foot, and no other intervention. We have violated the very sanctity of race car driving with these PUs that are dependent on automated algorithms that continuously remap the throttle pedal and deliver different power outs for the same pedal position. It's not race car driving. It's robo-race. If the car is going to automatically adjust it's power level by itself, why not install some LIDAR and have the cars drive themselves autonomously. There's no point having a driver in the loop. I am not in favor of robo-race, but current F1 is robo-race with added steps (a fussy, emotional driver interfering with the automated power delivery).

[Tommy Cookers]

... It should have never become "controversial" for the throttle pedal to have a single power output level that is scaled only by the position of the drivers foot, and no other intervention. We have violated the very sanctity of race car driving with these PUs that are dependent on automated algorithms that continuously remap the throttle pedal and deliver different power outs for the same pedal position. It's not race car driving. It's robo-race. ....

didn't this start in the 90s with 'drive-by-wire' ?
(reined in with the 2000s rules reducing but not eliminating the extent of engine 'torque-mapping')
rules that were built into 2014-2025 and their effect limited by the 120 kW MG limit

'Airbus rules'- vehicles designed by people who can't drive vs vehicles driven by people who can't design ?
it started in 1940 eg when 'they' changed the story of how planes were to be flown ....
or when the first understeering GP car appeared (Uhlenhaut's 1937 Mercedes-Benz)

and wow ! energy spent early on the straight is 'efficient'
Coventry Climax said average power is what counts (not peak power)
that's why they won a lot of races

[michl420]

This different output per pedal input comes with this flicked regulation in this year. In Melbourne full throttle was always full power (at the beginning of the full trottle phase).