Answers are grounded in this paper and cite their sources.
Autonomous UAV navigation in cluttered warehouses requires anticipation of dynamic and occluded hazards, explicit risk attribution, and recovery-aware action selection. CROWN-UAV integrates an object-centric encoder, a retrained action-conditioned dual world model, stochastic future hypotheses, a calibrated structured risk graph, candidate-level runtime filtering, discrete receding-horizon planning, and planner-grounded audit traces. A frozen, leakage-controlled revision campaign evaluated five implementations on 120 paired NVIDIA Isaac Sim 5.1.0 PhysX cases per method across three warehouse layouts, four hazard families, five disturbance conditions, and two independent scenario seeds. CROWN-Full achieved 59.2% success (95% CI 50.2–67.5%), identical to Greedy and slightly below MPPI (60.8%). Relative to Greedy, CROWN reduced near collisions from 51.7% to 30.8% (paired difference − 20.8% points, Holm-adjusted p = 0.00346) and increased mean minimum clearance by 0.241 m (adjusted p = 5.40e-9), while successful paths were 2.50 m longer and completion was 3.90 s slower on 53 jointly successful cases. At 4.5 s, the DualWorldModel improved UAV ADE (0.313 versus 1.314 m) and occupancy IoU (0.527 versus 0.392) relative to constant velocity, but object ADE was worse (0.772 versus 0.547 m). Full-loop p95 latency was 101.95 ms, with 150/15,170 deadline misses. The evidence supports an implementation-specific proximity-safety trade-off and auditable decision records, not higher task completion, universal predictor superiority, formal safety, or human-validated explainability.
No takes yet. Share an insight, caveat, or question.
Farkh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: