WorldFly: A World-Model-Based Vision-Language-Action Model for UAV Navigation

WorldFly introduces a world-model-based Vision-Language-Action framework that enables UAVs to navigate complex urban environments by predicting future states rather than relying solely on immediate observations. The system uses a dual-branch coupled flow matching mechanism to generate both video predictions and navigation actions, addressing critical limitations in dense urban scenarios with severe occlusions and sharp directional changes.

WorldFly represents a significant advancement in autonomous aerial navigation by addressing a fundamental limitation in current vision-language-action models: their inability to handle severe partial observability in challenging environments. Traditional VLA systems operate reactively, using historical observations to predict immediate actions, which fails when sharp turns or building occlusions obscure the path ahead. By integrating world models—systems designed to predict future environmental states—WorldFly enables UAVs to "imagine" upcoming scenarios, fundamentally improving decision-making under uncertainty.

The research emerged from observed failures of existing approaches in dense urban canyon environments where viewpoint transitions are drastic and continuous. Rather than accepting these limitations as inherent to the problem, the authors constructed a specialized benchmark specifically targeting these failure modes, enabling rigorous evaluation of spatial understanding in high-complexity scenarios.

The dual-branch coupled flow matching mechanism represents the technical innovation enabling this improvement. By jointly generating future video predictions alongside navigation actions, the framework creates an explicit feedback loop where spatial imagination directly constrains policy decisions. This architectural choice transforms the navigation problem from pure reactive control into informed planning.

The demonstrated performance improvements, particularly in unseen environments, validate that world model integration transfers knowledge more effectively than traditional approaches. For autonomous systems operating in GPS-denied urban environments, this advancement has direct practical implications for search-and-rescue operations, infrastructure inspection, and autonomous delivery systems. The work suggests world models should become a standard component in embodied AI systems operating in visually occluded environments.

• →WorldFly combines world models with vision-language-action frameworks to enable UAVs to predict future states and navigate occluded urban environments more effectively than reactive approaches.
• →The Urban Canyon Traversal Benchmark provides a rigorous evaluation framework for testing spatial understanding in scenarios with severe occlusions and drastic viewpoint transitions.
• →Coupled flow matching enables joint generation of video predictions and navigation actions, creating explicit spatial imagination to guide UAV policy decisions.
• →Performance gains are particularly pronounced in unseen environments, suggesting world model integration improves generalization beyond training distributions.
• →This advancement has practical implications for autonomous systems in GPS-denied urban scenarios including delivery, inspection, and search-and-rescue applications.