New research published on arXiv introduces two novel computational methods designed to improve planning efficiency for autonomous systems: the GeoWind2Plan urban wind prediction model for unmanned aerial vehicles (UAVs) and a risk-sensitive Partially Observable Markov Decision Process (POMDP) planning framework based on Conditional Value at Risk (CVaR). Both studies tackle a fundamental challenge in autonomous navigation: handling environmental uncertainty while ensuring safe, real-time decision-making.
GeoWind2Plan: Fast 3D Wind Field Inference for UAVs
According to the report detailing GeoWind2Plan (arXiv:2609.36056), complex 3D wind patterns around urban structures cause UAV energy consumption to depend heavily on localized wind fields rather than simple path length. Previously, high-fidelity computational fluid dynamics (CFD) simulations required several hours to days per boundary condition, making them impractical for UAV missions lasting only minutes or tens of minutes.
GeoWind2Plan overcomes this bottleneck using a local neural operator conditioned on 3D building geometry. The model predicts the wind field along a flight corridor in roughly 3 seconds, compared to approximately 8 hours for conventional CFD. When validated via CFD simulations, trajectories optimized by GeoWind2Plan reduced energy consumption by 6.9% in tailwinds, 12.7% in headwinds, and 4.5% in crosswinds compared to wind-agnostic planning, recovering between 75.0% and 87.9% of CFD-based theoretical energy savings.
Risk-Sensitive Online POMDP Planning via CVaR
Published concurrently, paper arXiv:2609.35874 proposes a method for risk-aware online POMDP planning. Conventional online POMDP solvers generally optimize expected cumulative costs, which can neglect catastrophic tail risks when belief distributions place weight in hazardous regions.
Rather than modifying the full trajectory value function—which typically requires custom, computationally heavy algorithms—the authors apply the CVaR risk metric directly to the immediate cost over belief distributions at each step. This formulation preserves the standard MDP structure, allowing any standard online POMDP planner to become risk-averse simply by updating its cost calculations. The authors also provide mathematical proofs demonstrating finite error bounds between the approximation model and the underlying POMDP.