This video details an ambitious project to train an AI to achieve a world record in TrackMania's iconic A01 level, surpassing both human and tool-assisted speedrun (TAS) benchmarks. Initially, the AI uses reinforcement learning to independently discover and refine advanced driving techniques like specific downhill jumps and the "speed drift" bug, eventually beating the human world record after receiving targeted "rewards" for a challenging drift. The project then escalates to challenge a meticulously crafted TAS record. Facing limitations in AI precision, the creator introduces an "auto-drift" tool and segmented run methodologies. Ultimately, an unexpected discovery of a subtle "hole" in the track geometry during brute-force TAS exploration allows for a novel jump. This new strategy, combined with AI segments, leads to a new, faster TAS record for A01, which is then further optimized by the TrackMania community. The video highlights the interplay between AI precision, human ingenuity, and collaborative effort in pushing the boundaries of video game performance.
Summarizing the AI's Journey to TrackMania A01 Perfection
0:02 Challenging the Human World Record: The creator trains an AI to beat Eddie's legendary world record on TrackMania's iconic A01 level, aiming to explore the true limits of the game.
1:19 Reinforcement Learning: The AI learns to drive through reinforcement learning, receiving rewards for progress, trying new actions, and iteratively improving its strategy through trial and error.
3:00 First Major Breakthrough (Downhill Jump): After thousands of attempts, the AI rediscovers a key technique in the downhill section: releasing the accelerator before a jump to land earlier and regain speed, later refining it with diagonal jumps.
4:19 Second Major Breakthrough (Speed Drift): The AI independently discovers the "speed drift," an advanced technique exploiting a game bug where braking at a specific angle causes faster acceleration, enabling it to surpass the creator's times and reach the world's top 100 leaderboards.
5:55 Equaling the Human Record: The AI, after 400 hours of training, comes within 0.05 seconds of Eddie's world record, but struggles to fully replicate a more complex second speed drift used by top players.
8:14 Targeted AI Assistance: The creator intervenes by giving the AI extra rewards for performing the difficult second speed drift, helping it adapt and eventually achieve times 0.04 seconds faster than any human.
10:12 New Goal: Beat the TAS Record: The AI's new objective becomes to surpass the A01 Tool Assisted Speedrun (TAS) record of 23.66 seconds, a theoretical ideal run meticulously crafted by players using external tools.
13:43 Introducing the "Auto-Drift" Tool: To overcome the AI's precision limitations (due to discrete action updates), the creator provides an "auto-drift" program that automatically holds optimal drift angles, allowing the AI to cede control for perfect execution.
15:44 Segmented Run Strategy: The AI adopts a TAS-inspired "segmented run" approach, where it optimizes individual track sections and then combines the best performances to achieve a flawless overall run.
18:53 The "Blue Bug" Myth: The creator explores the long-standing myth of a "blue bug," a theoretical glitch at track borders that could trigger an early jump in the downhill section, but extensive brute-force attempts fail to consistently reproduce it.
22:28 Unexpected Discovery: The "Hole Jump": During brute-force experimentation, the creator accidentally uncovers a previously unnoticed "hole" in the track (misaligned road pieces) that can be exploited for a powerful jump, offering a new path to beat the TAS.
25:43 New A01 Record Set: By combining the newly discovered "hole jump" with the AI's optimized segments, the creator successfully sets a new, faster TAS record for A01.
26:15 Community Collaboration & Further Optimization: The discovery is shared with the TrackMania TAS community, leading to a new team forming to further optimize the strategy, resulting in an even faster "masterpiece" TAS record of 23.62 seconds.
27:45 AI's Role & Future: The video concludes by noting that while the AI demonstrates superior precision, human creativity remains crucial for discovering entirely new strategies. The AI is now training on other complex maps.