January 1, 2026
Tour de Skynet
Cycling Game (Mini Neural Net Demo)
AI pedal power: robot cyclists learn to win while comments go full Tour de Drama
TLDR: New cycling sim lets identical riders with simple AI brains evolve race tactics by reading speed, slope, and battery, then mutating winners. The crowd loves watching draft monsters and sprinters emerge, while debating single-layer brains, clamoring for extra inputs, and joking about “AI doping”—because physics meets strategy is wildly fun.
Robot cyclists with tiny digital brains take to a bumpy track and learn to race. Each rider has the same body, but a mini brain reads simple cues—speed, power, battery, steepness ahead—and decides how hard to pedal. Winners get cloned or tweaked, losers vanish; press Space to fast‑forward evolution. Viewers are gleefully mashing reload for weirder hills and wilder tactics. One fan, ungreased0675, summed it up: “mini neural networks trying to find the best plan.” Meanwhile JKCalhoun is in full lab‑coat mode, asking if there’s “a single layer” and dissecting the clever two‑scale slope inputs (100m vs 1000m). The red/blue input glow is catnip for armchair coaches.
The drama: is smart pacing enough, or is physics the real boss? Some cheer the obvious rule—push uphill, chill downhill—others stan pure draft monsters and last‑second sprinters. A growing chorus wants more senses, like “less than 100…” meters hazard lights, while purists say the simple brain keeps it fair. Jokes fly about “AI doping,” “Skynet peloton,” and turning the anaerobic battery into “banana juice.” And yes, spectators proudly admit spamming Space to force evolution. The creator @ajddavison invited suggestions, and the crowd delivered—along with memes, debates, and a lot of sweaty virtual handlebars.
Key Points
- •Each rider is controlled by a small neural network using defined percepts to adjust power output per timestep.
- •After each race, the top five riders seed the next generation via exact copies and small weight mutations to evolve performance.
- •The simulation models cycling physics with 87 kg mass, CwA=0.32, Cr=0.004, and up to ~40% drafting benefit when following closely.
- •Physiology is modeled with a 250 W aerobic threshold, a 15,000 J W′ anaerobic battery, and a 750 W max sprint that declines with battery level.
- •Users can inspect controllers, switch views, force early evolution, and reload to reset riders and randomly generated terrain.