Learning Active Depth Sensing for UAV Navigation: Reliability Supervision and Staged Adaptation

Adaptive depth acquisition changes the observations used for UAV navigation. We present a staged framework that connects differentiable depth-reliability supervision to flight performance. An active-stereo surrogate supplies power, exposure, and gain targets for the least reliable image patches. Student-rollout relabeling extends supervision to learned-acquisition states, and independent visual stems preserve the camera mapping during flight adaptation. In the primary simulated aperture-crossing benchmark (60-step cap), success reaches 75.1% versus 66.7% for nominal fixed acquisition, while valid-depth fill rises from 72.1% to 97.5%. Under a separate ablation protocol (5 s timeout; 120-step cap), removing relabeling and retraining lowers success from 77.89% to 70.78%; sharing the visual stem lowers it to 74.30%. The frozen policy retains 75.33%–79.67% success across four partial-observation layouts. A bounded-depth-bias test yields 70.00% success at 0.10 m and 37.00% at 0.30 m, compared with its 74.67% zero-bias control. Supplementary D455 captures show a coverage gain after camera adaptation, and a separate PX4 hardware-in-the-loop test completes 44/60 crossings with synthetic depth. The results demonstrate navigation gains from the complete differentiable-sensing and flight-adaptation framework while quantifying its sensitivity to incorrect ranges.