Teaching drones to follow directions by predicting what comes next

Aerial vision-language navigation requires drones to follow natural-language instructions in 3D environments using closed-loop perception and action. WorldVLN frames this as prediction-driven control: the agent anticipates how the world will evolve, then acts accordingly. Rather than generating full video sequences, it uses a latent autoregressive backbone to predict short-horizon world-state transitions and decode them directly into waypoint actions, with newly observed information fed back into the loop after each action. A two-stage training approach first grounds video priors in navigation dynamics, then applies Action-aware GRPO—a novel RL method for autoregressive world-action models—to optimize actions through rollout consequences. On public outdoor and indoor benchmarks, WorldVLN outperforms existing baselines by 12%+ on success rate and shows stronger gains on harder cases. Code and demos are released.