Stuff I've put together,

KAIST Introduction to Reinforcement Learning · Semester-long team project

Korean 4-Ball Billiards RL

A continuous, deterministic, sparse-reward RL benchmark built on a from-scratch physics simulator for Korean 4-ball (sagu). Off-policy RL alone plateaus below 1 point/inning; a geometric aim constraint and carom features lift SAC to 6.460 points/inning; inference-time depth-2 lookahead using the simulator as its own verifier then chains up to 9,392 consecutive scoring shots at 99.8% per-shot success.