Quadrotor navigation policy transfers to physical drone without further tuning

Researchers report that a quadrotor navigation policy trained with explicit obstacle-avoidance constraints transferred to a physical drone without additional fine-tuning, according to an arXiv preprint. The policy also transferred to Gazebo/PX4 and generalized to denser environments outside its training distribution. The method trains through a differentiable simulator, using a clearance-based constraint and an aggregation function that emphasizes critical safety violations. This lets training specify a desired safety level directly rather than manually balancing a fixed penalty against task performance. The authors report improved safety and performance tradeoffs over tested baselines; their convergence proof depends on assumptions.