Answers are grounded in this paper and cite their sources.
Close-proximity quadrotor formations require disturbance compensation, but learned feedforward can have observer-dependent value. We examined a retrained neural force prior with zero ambient-wind inputs through complementary-filter analysis and a complete prior-observer-robust factorial. Five training seeds used grouped synthetic data, while independent tuning and 50 paired evaluation seeds separated component increments from tuned-method comparisons. Each comparison regime contained 5400 entries across six scenarios, with complete-period, gust/recovery, and peak error assessed jointly. The neural prior reduced period error by 0.13–0.98 mm in five complete matched families, but slightly increased peaks in three. Failed executions prevented complete-family inference for combined stress. Public two-vehicle records supplied a separate source-platform load-shape check with held-out recordings.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: