Thanks for watching!
Some additional details not explained in the video, which might help to better understand the irregularities observed:
- I didn't show all the AI inputs in the video to keep things simple. In reality it has access to more information, such as x,y,z velocity, velocity rates, roll-pitch-yaw rates, etc. But maybe it's still missing some crucial information, it's hard to know.
- The irregularities observed are not due to hardware or framerate issues. All the things I'm showing in the video are made with a tool called TMinterface. This tool allows me to inject action commands into the game at precise timings, in a way that is 100% repeatable. The same sequence of action on the same map will aways lead to the same outcome, even on a different computer.  It's completely deterministic. And it's used a lot for Tool Assisted Speedrun (TAS), you can find many examples of that on youtube.

I have a bit of extra footage that I couldn't fit into this long video. I plan to post some of this extra content on my Patreon in the next weeks. Any support there is a great motivation to keep making these videos :)
patreon.com/Yoshtm

I came here to see AI destroy humans on Trackmania, but I got a 37 minute essay on how pipes in Trackmania are an exercise in chaos theory instead. 10/10

PhD Student in Deep RL here. The behavior at the end seems mathematically chaotic and might lead you to the conclusion that you cannot predict deterministically the behavior of the car. However, that does not mean you can't improve the performance by a long shot with a few simple tricks.
You are doing model-free reinforcement learning, which basically means that you don't need to predict what is going to happen exactly (and extremely hard here), you just need to figure out what actions are the best in this situation. In most environments, this is actually much easier to do (and a reason why learning an accurate model is often harder than just straight up optimizing for performance with model-free).
Second problem is that you are using an RL algorithm and RL assumes you are in a Markov Devision Process, ie that you have full observability of the world, ie that the probability of taking the next action ONLY depends on the current information that you have, and not the past, ie that adding information about the past does NOT help ou make better decisions. HOWEVER, you are NOT in an MDP (Markov Devision Process) here since you lack critical information about the state to take decision, as many pointed out.

In practice, what you can do will most certainly improve performance:
- either manually add all the missing info (speed, rotations, ...)
- or more simply, create a concatenated vector of  the previous N  past states and inputs and use that as input to your NN (LSTM/RNN works too but is more complex and often not necessary when we are dealing with a few steps of history).
- Use more layers (deeper network  = more complex function can be learned by the network)
- Sticky/random actions: introduce manually randomness in the action of the agent. For example repeat randomly the past action with a small probability. Do NOT decrease this probability with time. What this means is that the agent will have to learn the whole time in a now stochastic (ie random) environment and will have no choice but to cope with the inherent randomness that you added, making it largely more robust than an deterministic agent. This is a common flaw of fully deterministic agent in these kind of environment btw.
- Look at extensions like Max Entropy RL (Soft Actor Critic for instance), where you both maximize reward and entropy of the policy. Some papers proved that it makes the learned policy more robust to out-Of-Distribution perturbations, ie perturbations that were never seen during training. In your case, this will help the agent lean "recovering behaviors" to recover from deviations caused by physics bugs/randomness/whatever.
- Max Entropy RL will also help with exploration a lot, which might help the AI be more creative in finding solutions.
- Look at other tricks that improve performance "for free". I don't know how much you implemented already, but a basic start is looking at Rainbow DQN, PPO.... But you are already using it probably..?

ah yes, my favorite insaneley difficult map. "One hella long pipe" lmao

28:50 Bro made me question his patience level 💀

So you’re saying my world record fails are actually the butterfly effect’s fault…

I love how the AI on the 65km pipe track looks like it's just blissfully skipping along while the human records look like they're terrified. it's so silly.

I like how yosh's attitude towards the learning model went from "We shouldn't have to sacrifice consistency for speed" to "Fuck it we ball"

love the ai dancing on the pipe 4:21

The song that played when you showed Wirtual's WR was just peak video making lol

Yosh: "I am going to lock your breaks and acceleration so you cant slow down"
AI: "Pfft check this out"

Thanks for using my map :D
I have heard people saying that Trackmania is a deterministic game hundreds of times. I think from now on I will prefer to use the word "chaotic" :D

10:54 the graveyards of fallen AIs

Roboticist here. Some regions of configuration space are more chaotic than others; the demo you did where you perturbed a car that was about to fall off demonstrated a smooth increase in success rate as you moved the perturbation further away from the point at which outcomes diverged, suggesting that the failure was in a fairly narrow chaotic region and moving out of that region made the system more stable and enabled more consistent success. This is what unnamed did to complete the jumps track: Found a strategy that consistently avoided highly-divergent regions of configuration space.

One problem is that, if the display of the inputs is representative, the AI doesn't have the right sensors. Specifically, the AI has its position and speed but not its _velocity_; it doesn't know how much of its speedo reading is in each component of down the track vs. across the track vs. away from the track. More critically, it also doesn't have roll-pitch-yaw rates. It only knows its current orientation, not the rate at which its orientation is changing.

I suspect that the car also needs to know the difference between a simple pipe elbow, a tee, and a cross. Even if the extra geometry isn't in the direction of its travel it'll almost certainly still interact.

Finally, consider increasing the numerical precision of your network. This is one of those cases where running floats vs. doubles might actually matter.

22:48 "THE AI GOT THIS RUN" omg even the song, that's great man

28:50 “Now let’s have the AI drive upside down” had me rolling. For a hot second I really thought you were about to blow my mind once again 🤣

i love the ,butterfly effect‘ detail in 17:27 ^^

I loved how it just randomly turned to a Wirtual documentary for a while at 21:55, even the Norwegian flag is here.

yooo such an honor that you used my map for the 3rd part ♥really enjoying your AI videos :) keep up the great work!

Your analysis about chaos theory is exactly what is going on. And TrackMania's chaotic physics are very familiar to me from another TrackMania project. The best way to control that chaos is by going slower, as the deviations caused by the bugs are less severe at lower speeds. Perhaps giving less reward for speed and more reward by having a high chance of getting far would help with consistency.