The 2001 Nokia classic, played by PPO agents trained on a C rewrite of the original
The original 2001 Nokia Bounce — all 11 levels, reverse-engineered from the JAR — running entirely in your browser via WebAssembly. The red ball is controlled by a separate neural network policy per level: the Level 1 policy clears Level 1, then the page switches to the Level 2 policy for Level 2, and wraps back to the start. Levels 3–11 are not trained yet.
Environment: C rewrite of the JAR game logic, 164-dim observations.
Level 1 policy: 998M timesteps, 99.9% clear rate. Level 2 policy: 524M timesteps, 99.8% clear rate.
PPO, trained on an RTX 5070.