Goaltargetnet: Goal Prediction and Tracking for F1tenth Autonomous Sim Racing League With Convolutional Neural Networks

In this paper, we introduce GoalTargetNet, a convolutional neural network (CNN) architecture trained for autonomous racing in the F1TENTH Sim Racing League. Unlike traditional end-to-end methods leveraging deep learning to control autonomous racing cars that directly predict control signals for speed and steering, our approach replaces localization and motion planning with a goal-targeting network. Paired with a control component, the whole system uses a hybrid approach to solve the autonomous racing problem. The GoalTargetNet model is trained on LiDAR data to predict a target goal and recommended speed, balancing performance and safety. We demonstrate that our approach achieves competitive real-time performance while maintaining a lightweight model suitable for deployment on physical autonomous racing cars. Our results suggest that GoalTargetNet offers a promising alternative to both classical and fully end-to-end autonomous racing strategies.