PPO · autonomous control · TU Dresden project

Reinforcement Learning Racing Agent

I implemented a PPO-based autonomous agent for 2D racing using reward shaping, curriculum learning, and sensor-based perception.

The challenge

Train a racing agent to progress from basic control to stable navigation without relying on a single brittle reward signal.

I combined PPO training with reward shaping, curriculum learning, sensor-based state information, and iterative training strategies.

Reward design and curriculum progression were treated as engineering decisions, not hidden training details.

This TU Dresden project is presented separately from a later collaborative racing project because their relationship has not been confirmed.

A reproducible applied-RL project focused on learning stability and autonomous control behaviour.