This approach, at the technical level, involves:

The method integrates calibrated camera intrinsics and extrinsics for distortion correction and frame alignment, motor gain and trim calibration for odometry consistency, and ROS-based perception nodes for lane filtering, stop line detection, obstacle recognition, and AprilTag-based pose estimation. Control nodes implement PID regulators for velocity and steering, parameterized turning primitives, and synchronized execution through a finite state machine that coordinates lane following, intersection stopping, turning maneuvers, and recovery states. Sensor fusion combines camera streams, encoder feedback, and ToF measurements for robust decision inputs.

Quantized Qwen 2.5 vision-language models were deployed with llama.cpp, configured with reduced context window and batch size to match GPU memory limits. The models were evaluated for trajectory planning and visual reasoning tasks, with both 7B and 3B variants tested under quantization schemes. Integration required Docker containerization for portability and ROSBridge for monitoring and remote interaction.

Challenges included GPU memory capacity restricting larger model execution, inference latency exceeding 100 ms control cycle requirements, CUDA feature mismatches across builds, and instability in container runtimes on the NVIDIA Jetson Nano platform. These issues necessitated systematic parameter tuning of controllers, quantization of VLMs to GGUF formats, pruning strategies to reduce computation load, and hybrid offloading of visual reasoning to external compute nodes while maintaining low-level perception and control locally. Additional constraints involved balancing message-passing overhead in ROS, synchronization delays between perception and control nodes, and variability in inference reproducibility across different hardware builds.