Problem

Autonomous vehicle systems in controlled environments, particularly for racing, face significant challenges in integrating various subsystems—perception, localization, path planning, and control—for real-time navigation and decision-making. Developing robust and modular architectures that effectively combine these elements is critical for achieving precise operation in constrained, competitive scenarios.

Method

The proposed system utilizes a modular architecture with four main components: Perception, Localization and Mapping, Trajectory Planning, and Control. The hardware setup includes a dual-computer system with an NVIDIA Jetson Orin NX for high-performance computing and a separate in-car PC for CAN communication and ROS interfacing. The software stack is based on ROS2 Galactic and employs Docker for virtualization.

The Perception module uses a ZED 2i stereo camera for image capture and integrates its Neural mode for AI-powered depth calculation. Object detection for track cones is performed by a YOLOv11s model, trained on the FSOCO dataset. An Intersection-Over-Union (IOU) tracker is used to maintain cone identities across frames, and 2D positioning is derived using a distortion-free projective transformation based on a pinhole camera model.

The Localization and Mapping module combines centimeter-level GNSS accuracy from a simpleRTK2B receiver with IMU data using an Extended Kalman Filter (EKF) and a Kinematic Bicycle Model. This fusion aims for high-precision positioning. The mapping node continuously stores and updates cone and car positions, averaging cone detections inversely proportional to distance to reduce depth errors. For missions with known tracks, a Global Map Alignment module can load a pre-defined map and align sensor data using a transformation matrix.

The Path Planning node calculates an optimal racing line by interpolating blue and yellow cones using 2D cubic splines to define track boundaries. The middle points of these boundaries form the racing line, which is then discretized with assigned target speeds. The algorithm adapts its behavior for initial "discovery" laps versus "timed" laps, activating global path planning for closed loops. For non-closed circuits like skidpads, an alternative pipeline is used to align the ego vehicle with a pre-defined global map.

The Control module directs the vehicle along the planned path using a Regulated Pure Pursuit algorithm. This algorithm dynamically adjusts a lookahead distance to determine the optimal steering angle and calculates acceleration based on the difference between current and target speeds to avoid overshooting. A master node oversees the entire system, managing node activations, internal flags, and triggering the Emergency Braking System (EBS) if unexpected states are detected.

Results

The architecture was evaluated with individual modules tested using real-world data and the end-to-end system tested in a simulated environment that replicated observed sensor error distributions.

For object detection, the YOLOv11s model achieved a 0.93 mAP@0.5 on the test set after hyperparameter optimization using a genetic algorithm. Depth calculation tests revealed an exponential increase in error with distance, reaching a median error of 0.5 meters at an 8-meter distance, with a tendency to place objects closer. RTK-GNSS data demonstrated improved accuracy over raw GNSS, reducing error to approximately 12 centimeters and experiencing fewer signal losses. EKF fusion was shown in previous work to improve localization accuracy by about 32% compared to raw sensor data. The car control algorithms were successfully deployed and tested within the simulated environment, taking into account the observed sensor error distributions.

Implications

This modular architecture provides a robust and integrated framework for high-precision autonomous driving in controlled racing environments. The system effectively combines state-of-the-art computer vision, advanced localization techniques, adaptive path planning, and regulated control, demonstrating a practical approach to real-time navigation. The detailed consideration and mitigation of sensor errors, alongside adaptable mission-specific functionalities like global map alignment and discovery laps, contribute significantly to the system's reliability and performance in competitive or experimental autonomous driving scenarios.