Second Place Solution in Fall 2025 Simulation Racing Series

Additional improvements in completion time could not be achieved through trial runs. An automated solution was necessary for further improvement. A Distributed Evolutionary learning Algorithm (implemented using the deap library) that sits outside the simulator and reasons entirely from observed outcomes was built. Each run starts from a clean slate by fully restarting CARLA and then launching competition_runner.py again before collecting any data. This reduced run to run variability caused by lingering physics state timing drift or caching behavior and makes comparisons between configurations meaningful rather than noisy.

After each run is completed, the program reads the structured results produced by the competition framework. These include section completion times per lap, total lap times, whether CAS was activated, and whether a crash has occurred. Instead of trying to predict behavior inside the simulator the AI treats these results as evidence of how a specific configuration behaved under realistic variability. It evaluates performance by aggregating time and event-based signals rather than assuming that faster runs always mean better. This allows it to recognize that some settings work well on average even if a single run is imperfect.

The only control variable the AI is allowed to modify is µ. This variable directly scales aggressiveness through the velocity rule given by equation (2). Higher value for µ increases corner entry and exit speed while a lower value makes the car more conservative. Separate µ values for each section were explicitly defined because tire grip conditions, behavior, and stability change over a run. Applying one value across all laps produced worse results in practice. All µ values are clamped between 2 and 4 which was the range determined empirically as the safe and meaningful operating range. Values below 2 were consistently too slow and values above 4 produced frequent CAS activation and crashes.

The reward model is numeric and intentionally asymmetric to reflect risk. Completing a section without incidents (crash) adds positive 1 reward. Completing a full lap adds positive 10 reward. If a lap time improves relative to the best previously observed time for that lap it adds an additional positive 5 reward. Each CAS activation adds negative 5 reward because it indicates that the car exceeded stable limits even if it did not crash. A crash adds negative 50 reward and immediately terminates learning for that run since post-crash data is not representative of the target performance during a race. If the total run time is slower than the previous reference, an additional negative 10 reward is applied even if no crash occurred.

Once a run ends the AI compares the total reward and the lap and section times against a rolling history of prior runs rather than a single baseline. This matters because CARLA exhibits natural stochasticity and judging performance off one comparison leads to overfitting. If the configuration shows statistically better behavior across time, faster laps fewer CAS events, or better reward consistency the AI nudges the corresponding lap µ upward by a small, fixed step of 0.05. If the performance degrades or CAS frequency rises the AI nudges µ downward. These updates are deliberately small, so the system explores locally around known good regions rather than jumping into unstable regimes. Because µ is adjusted independently per lap the system can learn for example that a higher µ works well on Lap 1 but needs to be reduced later once the car becomes more difficult to control.

The learning process ran entirely on an RTX 5070 TI GPU. Although there is no heavy neural network the GPU is used for fast numerical aggregation and comparison across many runs which allows high iteration counts without becoming compute bound. Over repeated restarts and evaluations the AI converges toward configurations that are fast but stable under variability rather than simply optimizing a single best-case run.

The values in Table IV show the values µ obtained through the application of this lightweight learning loop.