Loading the latest organizer resource…
TRACK DETAILS / TWO DIVISIONS
Build every decision with logic, or train a neural network to make the call. Explore the methods, boundaries and deliverables for each division.
Compare the tracksSPELLCRAFT VS NEURALCRAFT
Follow two approaches through one corner. See how a programmed controller and a neural policy can turn observations into driving decisions.
Read speed, grip & road preview.
Compare speed with the braking target.
Correct heading & follow the reference line.
Unwind steering & increase throttle.
The highlighted stage shows which programmed decision is being inspected.
EXPLICIT RULES
Explore SPELLCRAFTSwipe to inspect all layers →
A tiny network computes these actions in your browser. Its fixed weights are hand-authored for illustration, not trained.
ILLUSTRATED NEURAL POLICY
Explore NEURALCRAFTReady. Play the demo or choose a phase.
Both drivers receive the same road observations.
Educational illustration: cars follow a shared scripted path, not a physics simulation. Neuralcraft actions are computed by a tiny network with hand-authored weights; Spellcraft values follow the rule storyboard. Neither represents a trained competition entry or a performance ranking.
01 / CHOOSE YOUR APPROACH
Both divisions use equivalent vehicles, observations and evaluation scenarios. Your runtime decision-making method determines the division; the tools you use to write code do not.
02 / LOGIC DRIVERS TRACK
ENGINEER EVERY DECISION.Build a driver from rules, equations, optimization and state machines you design. No trained model may participate in the runtime decision path.
| Allowed approach | Example use |
|---|---|
| PID, Pure Pursuit, Stanley | Correct speed, heading and cross-track error. |
| Model Predictive Control | Optimize a local trajectory and actuator commands. |
| State machines and handwritten fuzzy rules | Switch between cornering, cruising and legal recovery. |
| Offline parameter search, including genetic algorithms | Tune a classical controller; the race-time driver remains classical. |
A trained policy disguised as a lookup table is not a classical controller. Generative coding assistance is permitted in the proposal when disclosed; it does not make a driver Neuralcraft.
03 / NEURAL DRIVERS TRACK
TRAIN THE MIND BEHIND THE WHEEL.A trained neural network must meaningfully affect driving. It can return actions directly or produce a speed target, trajectory or manoeuvre that a classical controller executes.
| Allowed approach | Example use |
|---|---|
| Deep reinforcement learning | Learn a neural policy from simulator interactions. |
| Imitation learning | Learn from legitimately sourced demonstrations. |
| Neuroevolution | Use a genetic algorithm to optimize neural weights or architecture. |
| Generative driving | Generate a trajectory or driving plan used by the car. |
| Hybrid neural control | A neural planner supplies targets to a classical controller. |
Tabular RL, genetic algorithms that only tune PID, and an LLM that only explains a replay do not meet the neural requirement. External commercial inference APIs are outside the proposed first-season championship.
04 / SHARED SUBMISSION
The organizer will publish final packaging requirements and deadlines. Try the controller contract preview →
