Neuron Platformer RL
Reinforcement learning demos usually run inside somebody else's environment, which hides the hard parts: reward design, observation design, and whether the task is solvable at all. Here the whole stack is built, including a level generator whose every gap is capped by the exact jump physics and an audit that re-proves solvability across 1,200 seeds. Then the real question: an agent reading raw 84x84 pixels, with no coordinates, velocities or object lists, beats the hand-crafted state-vector baseline 80% to 27% on 200 held-out levels, and a difficulty curriculum carries it to 52% medium and 55% hard. Grad-CAM shows where it looks: platform edges before a jump, the gap, the enemy. A fourth phase replaces the frame stack with an LSTM reading one frame at a time: recurrence does learn the game, but lands behind the frame stack on every tier - 29.5% medium against the curriculum's 52%. Getting these numbers honest took three extra experiments: doubling the LSTM's budget refuted the write-up's own excuse (typical performance rose, the peak did not), a noise analysis showed a 30-episode evaluation is worth about plus or minus 14 points, and a 200-seed re-measurement of all four models replaced every headline figure - the state baseline's published 50% was really 27%. The write-up was rewritten rather than left flattering.
pixel agent 80%
state baseline 27%
impossible transitions 0 of 11,619
PythonPyTorchStable-Baselines3GymnasiumGrad-CAM