Safe-VLA: Bridging Linguistic Reasoning and Safe Trajectory Planning for Autonomous Driving

Research Project, 2026 – 2030

Purpose and goal
Safe-VLA explores vision-language-action models to enhance road safety by improving the reliability and efficiency of autonomous vehicles´ decision-making in complex real-world scenarios. By integrating language-based reasoning with safety-aware trajectory planning, Safe-VLA enables a deeper understanding of traffic contexts and generates safe vehicle trajectories. This ultimately helps AD/ADAS systems reduce accidents, injuries, and fatalities.

Expected effects and result
Safe-VLA project will deliver VLA-ready datasets, evaluation frameworks, safety-aware trajectory decoders, reasoning–action alignment methods, and efficient VLA architectures. Expected impacts include fewer unsafe trajectory decisions, reduced hallucinated reasoning, improved handling of corner cases, and increased safety in automated vehicles. The project will publish the scientific research papers, train a PhD graduate, and strengthen Sweden’s scientific leadership in safe autonomous driving.

Planned approach and implementation
The research builds on state-of-the-art methods while advancing beyond the SOTA across 5 defined work packages: dataset curation, evaluation, trajectory decoder, reason-action alignment, and efficiency.