What Kauã demonstrated

Built a DeepRacer reward function end to end and validated it with structured experiments. The final design used track-width-normalized center distance, a steering-aware speed term, and a floored on-track reward, then the student tested variants, measured noise, and correctly diagnosed an unconditional speed bonus as the exploit. They also explained why the function generalized poorly when a finish bonus was tuned to the practice circuit’s step count and proposed a concrete fix.

• Leakage-Free Reward Design
• Honest Measurement and Noise Awareness
• Reward-Hack Diagnosis
• Generalization Reasoning Across Tracks