Something I thought of a while back, but this video added to it - interesting thing at 0:56 It seems like car in front had jumped, but that's just the multiplayer thing - last time info from that player's machine got through, his car was going uphill, so it continued that way until new information was received over network Seems like prediction code takes gravity into account, but not downforce. So here's a suggestion - I think it would be a nice thing to have AI physics applied to other cars in multiplayer. Note I'm talking about "AI physics", not AI itself. It would add a lot of immersion to multiplayer experience and prevent situations like in that video. You would have all the suspension movement and other stuff visible much better than by just visualizing and interpolating what you got over network. Just imagine better reactions of car going over curb in multiplayer. The difficulty here is that it would require information on other car's setup, so it would have to be sent over network. At least wing and suspension settings would have to be included. Engine and tire simulation could be skipped (except for collision detection). Just adding these bumps and impact forces, translating them through suspension and applying downforce should make a lot of difference for immersion in multiplayer.
I read and tried to understand this idea, but did not fully. If you receive 2 position samples of another client between xxx milliseconds, then the prediction code calculates a likely position for the car between these 2 points (not by using gravity or anything as far as I know, just by looking at the car's previous path and applying a smart algorithm). Now in between these 2 points, there is no real information telling where the car is. It can't be made to look like AI because there is not enough information to do calculations on - there is just two measurement points. If the prediction is made too aggressive, then things go even more wrong, which is what we had with some build last winter. The prediction would "overshoot" and the car would jitter and warp from place to place between each position data received.
Are you sure your not describing a hindsight prediction model. (No such thing) Remember you only receive two points you predict a third (the one that hasn't happened) until you get the next one. Then a correction can occur. Just easier to explain it that way. You were describing this but referencing it in past tense. I believe to keep the cars smooth the prediction model lets the cars position in part by this and on part by actual data. When it went wrong was when it was trying to average the two with a bug in the calculation of positioning? That is how I read that issue. I think what k Szczech is saying is that, that prediction part for smoothing happens in a z axis that is ignorant to all the parameters it should be. So a rises and drops look at trajectory only and do not apply basic physics to the average position. He is asking that simple physics are applied to the cars prediction or so that it doesn't do anything to wild like take off? But only he can correct us
Yes, I think I understood what he meant now, that locally on each client's PC, rF2 would calculate gaps in prediction with AI physics model. It sounds very complicated to me moving from simple z-axis model to something like this, but why not.
