AI Drones Now Obey Safety Rules and User Commands Together
Safer delivery drones and robots that follow instructions without breaking rules.
Logic-VLA is a vision-language-action model that takes formal Signal Temporal Logic (STL) requirements at inference time and adapts its behavior to satisfy them while following natural-language task instructions. Trained with STL-conditioned fine-tuning and trajectory-level preference optimization, it improved STL satisfaction by 24.8 to 40.7 percentage points over an STL-blind base policy in closed-loop quadcopter navigation simulations—while reducing nominal task success by at most 1.8 points. The same model handled varied formal requirements without needing a separate policy for each specification, including STL formulas it had never seen during training.
- The new AI lets drones understand both human commands and formal safety rules at the same time.
- In simulations, safety rule compliance improved by up to 40.7 percentage points with minimal loss in task success.
- One flexible AI can adapt to different safety requirements, meaning no need for separate software for each rule.
Why It Matters
Safer, rule-following robots mean delivery drones and automated helpers can work in crowded spaces without risky mistakes.