New AI drone planner cuts obstacle violations from 40% to 9.6%

Autonomous drone navigation traditionally splits perception, mapping, and planning into separate stages, propagating errors and requiring environment-specific retuning. End-to-end generative models avoid those interfaces but fail on glass, mirrors, and overexposed surfaces — they treat unreliable observations as valid evidence. A new paper from Skoltech researchers introduces a reliability-aware diffusion planner for 3D UAV flight. The system conditions trajectory generation on both the raw observation and a scene-level reliability heatmap produced by a lightweight network. This network distills open-vocabulary reasoning from a vision-language model, running within real-time planning budgets. The heatmap marks where perception cannot be trusted, and the denoising process is steered by a differentiable two-stage ESDF cost that treats physical obstacles and virtual obstacles (from unreliable regions) equally.

In simulation and on a real quadrotor, the planner reduced the obstacle-violation rate from 40.3% (state-of-the-art diffusion baseline) to 9.6%. Mean reliability of traversed regions jumped from 0.588 to 0.925. Ablating the reliability term alone dropped mean reliability from 0.898 to 0.783, confirming its decisive role. The distilled VLM runs up to 2× faster than the full model, making it practical for real-time deployment. The approach generalizes to unseen environments without retraining, addressing a critical safety gap in end-to-end drone navigation. Future work may extend the reliability-aware framework to other autonomous systems like ground robots or underwater vehicles.