AI just Broke Trackmania's most Legendary Record

The video introduces the goal: training an AI to challenge the A01 Trackmania world record. Initially, the AI drives randomly. Through reinforcement learning, it's rewarded for progress, learning to develop its own driving strategy through trial and error. After hundreds of attempts, it starts to make progress, seeing what the AI 'sees' as numerical data about the car's state.

After 3,000 attempts, the AI reaches a decent level. Its first major breakthrough is in the 'Downhill' section, where it learns to release the accelerator before the jump to land earlier and regain speed sooner. It refines this by jumping diagonally, effectively rediscovering a common player technique and eventually beating the developer's target time.

The AI makes its second major discovery: using the brake in the final corner to initiate a speed-drift. This advanced technique, when executed at a specific angle, exploits a game bug to accelerate the car faster, allowing the AI to close the gap on experienced players and reach the pace of the world's top 100 leaderboard.

The AI spends 400 equivalent hours on A01, getting within 5 hundredths of the world record. It mirrors the world record holder Eddie's lines, except for one critical speed-drift. Despite not adopting this second drift, the AI continues to improve, eventually equaling Eddie's record multiple times.

The AI struggles to learn the difficult second speed-drift due to extreme precision requirements at lower speeds. To help it, the creator gives the AI extra rewards for drifting in that section. The AI adapts, attempts the new drift, and despite initial slowness, refines its technique to eventually crush the human world record by 4 hundredths of a second.

Having surpassed human records, the AI now targets a new opponent: the Tool-Assisted Speedrun (TAS). This theoretical ideal run, created by players using external tools to design a perfect sequence of actions, stands at 23.66 seconds. The video focuses on the TAS in the classic category, explaining why other extremely glitchy TAS runs are excluded.

The AI loses speed compared to the TAS, especially in speed-drifts, due to its inability to hold the optimal drift angle precisely. This is attributed to technical limitations in the AI's action updates. To overcome this, a program is developed for the AI to use called 'auto drift,' which automatically follows the optimal angle with extreme precision, allowing the AI to achieve better and more consistent speed-drifts.

The AI still struggles to consistently achieve its full potential in a single run. Inspired by TAS methodology, the creator implements segmented runs. The AI drives small sections of the track multiple times, and the best segments are combined, allowing the AI to perform its best on every part of the track and get closer to the TAS.

Despite improvements, the AI still trails the TAS. The video introduces the 'blue-bug' myth: a theorized strategy involving triggering a car jump by crossing blue borders. This bug, if exploitable in the Downhill section, could eliminate the need to slow down, potentially leading to a faster time than the TAS. Many attempts to trigger it failed, but the creator becomes obsessed with the idea.

The creator attempts to find the blue-bug using a brute-force method by making small random changes to actions and replaying them. While searching, an unexpected jump is found. Upon closer inspection, it's revealed to be a tiny misalignment, a 'hole,' in the track's road pieces, a detail previously unnoticed. This discovery offers a new path to beat the TAS.

The creator focuses on exploiting the newly found hole. Through further brute-forcing from the point of the jump, they progressively achieve higher jumps, lifting the car to the jump line. Combining this with driving the rest of the track and calling the AI back to complete the run, a new limit for A01 is set, definitively beating the TAS.

The discovery is shared with the Trackmania TAS community, leading to a new team further refining the strategy and creating an even faster TAS. The video concludes by noting that records are always broken in Trackmania, and while the AI excels in precision, human creativity remains unparalleled. The AI is now training on other tracks in the campaign, aiming to compete with humans on more complex maps.