drivefusion

Autonomous-driving data pipeline for DriveFusion project built on the CARLA Simulator, generating cleaned multi-modal sensor data and VQA annotations for training vision-language action models.

Data preprocessing pipeline for the DriveFusionQA. It converts multiple autonomous-driving QA datasets into unified LLaMA and LLaVA-style instruction formats, with modular dataset preprocessors, JSON creators, and validation tools to support training and evaluation of vision-language models.

Autonomous vehicle deployment system for DriveFusion combining ROS-based control with AI vision powered by the Qwen 2.5 vision-language model, enabling real-time driving, single-shot decisions, and video/image processing.

Evaluation framework for the DriveFusion DriveFusionQA vision-language model, benchmarking Q&A performance on driving datasets using metrics like Lingo-Judge, BLEU, and BERTScore.