⛸️ 3D Proximal Policy Optimization: Clipped Surrogate, GAE and Policy Drift on a Slippery Ice Pond
Two Blender-built ice ponds with thin-ice holes and a fishing hut: a skater learns with PPO and its clipped surrogate objective while a twin learns with the same update but no clip; compare exact expected return, clipped-sample fraction, KL drift and policy entropy, backed by policy-gradient, GAE and Monte Carlo checks.
⚙ Under the hood
Two Blender-built ice ponds with thin-ice holes and a fishing hut: a skater learns with PPO and its clipped surrogate objective while a twin learns with the same update but no clip; compare exact expected return, clipped-sample fraction, KL drift and policy entropy, backed by policy-gradient, GAE and Monte Carlo checks.
Proximal Policy OptimizationReinforcement LearningClipped Surrogate ObjectiveGeneralized Advantage EstimationPolicy GradientTrust RegionMachine Learning
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install