Staged GT3 Setup Optimization with Setup-Conditioned Telemetry Response Modeling in Simulation

Engineering notes

Key topics: autonomous driving, prediction, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Optimizing a high-fidelity GT3 race car setup is a serious dimensional, nonlinear problem in which small changes to mechanical parameters can affect lap time, handling balance, and vehicle stability. Existing motorsport AI studies largely emphasize racing line optimization, autonomous control, race strategy, or offline vehicle dynamics estimation, while the mechanical setup layer is often treated as fixed or tuned manually. This paper presents a staged simulator-based setup optimization framework augmented with setup-conditioned telemetry response modeling. Using the virtual BMW Z4 GT3 vehicle model implemented within the Assetto Corsa (v1.16.4) simulation environment as a controlled GT3 test platform, 134 setup configurations were evaluated at the Red Bull Ring under a fixed simulator AI driving policy. The staged search improved the best lap time from 91.430 s to 91.040 s, corresponding to a 0.390 s reduction. To move beyond a single aggregate lap-time claim, the full telemetry corpus was processed into 585 stable laps and 29,250 track-position segment samples. A setup-conditioned LightGBM model was trained to predict segment time and local vehicle response metrics from setup parameters and segment context, using five-fold GroupKFold validation by telemetry file to avoid random row leakage. The setup-conditioned segment model reconstructed held-out file-level lap time with 0.223 s mean absolute error and Spearman correlation of 0.961, outperforming a setup-only model at 0.288 s, a track-only segment model at 0.687 s, and a shuffled-setup placebo at 0.776 s. The same setup-conditioned model also improved the prediction of segment-level speed, slip angle, tire load spread, rake (defined here as rear-front ride height difference), tire temperature, yaw response, and lateral acceleration. These results show that high-frequency telemetry can support not only staged setup search, but also quantifiable learning of where and how setup changes alter vehicle behavior around the lap.

Links and sources

Comments