Better Than Bots

Video Demo

About this project

Duck-3 is a Unity 6 game where a player competes head-to-head against trained ML-Agents models across three mini-games: 1. Pendulum: Balance an inverted pendulum by sliding a cart. The ML agent dynamically tunes PID controller gains (Kp/Ki/Kd) each step based on the rod's angle and velocity. 2. Lunar Landing: Guide a rocket to land upright. The agent does the same adaptive PID gain scheduling, but for rotational stabilization during a free-fall descent. 3. Projectile Catch: Position a bucket to catch a launched ball before it fires. The agent uses observations of the prepared launch angle and power to pre-aim; the player uses a UI slider. Each game runs for a configurable timer (default 30 seconds), managed by a persistent GameFlowController on the EventSystem that sequences the scenes automatically. Scores for both player and agent accumulate in a persistent GameScoreManager, and at the end, a Stats Menu shows a side-by-side comparison table of continuous reward scores for Pendulum and Lunar Landing, catch ratio (e.g., 2/3) for Projectile Catch.

Gallery