Bridging Perception and Reasoning: Dual-Pipeline Neuro-Symbolic Landing for UAVs in Cluttered Environments

TL;DR

NeuroSymLand tackles UAV landing in cluttered, map-poor environments by pairing an offline LLM-driven Scallop rule-synthesis stage with an online INT8-quantized perception backbone that builds a Probabilistic Semantic Scene Graph (PSSG). Executed rules form a white-box probabilistic first-order logic safety gate over the scene graph, yielding auditable go/no-go decisions with mission-conditioned ranking and provenance traces. The approach delivers improved accuracy, robustness to covariate shift, and edge-efficient operation, while offering interpretability, verifiability, and certifiability for safety-critical UAV deployments. Experimental results in SIL and HIL show superior SLZ quality, robustness, and runtime efficiency, enabling safe emergency, rescue, and delivery missions on resource-constrained hardware.

Abstract

Autonomous landing in unstructured (cluttered, uneven, and map-poor) environments is a core requirement for Unmanned Aerial Vehicles (UAVs), yet purely vision-based or deep learning models often falter under covariate shift and provide limited interpretability. We propose NeuroSymLand, a neuro-symbolic framework that tightly couples two complementary pipelines: (i) an offline pipeline, where Large Language Models (LLMs) and human-in-the-loop refinement synthesize Scallop code from diverse landing scenarios, distilling generalizable and verifiable symbolic knowledge; and (ii) an online pipeline, where a compact foundation-based semantic segmentation model generates probabilistic Scallop facts that are composed into semantic scene graphs for real-time deductive reasoning. This design combines the perceptual strengths of lightweight foundation models with the interpretability and verifiability of symbolic reasoning. Node attributes (e.g., flatness, area) and edge relations (adjacency, containment, proximity) are computed with geometric routines rather than learned, avoiding the data dependence and latency of train-time graph builders. The resulting Scallop program encodes landing principles (avoid water and obstacles; prefer large, flat, accessible regions) and yields calibrated safety scores with ranked Regions of Interest (ROIs) and human-readable justifications. Extensive evaluations across datasets, diverse simulation maps, and real UAV hardware show that NeuroSymLand achieves higher accuracy, stronger robustness to covariate shift, and superior efficiency compared with state-of-the-art baselines, while advancing UAV safety and reliability in emergency response, surveillance, and delivery missions.

Figures (6)

• Figure 1: NeuroSymLand pipeline. The left is online inference phase : SegFormer with lightweight post-processing (contours, distances, flatness/flow) extracts candidate regions. Cues are lifted to a probabilistic semantic scene graph; a Scallop-based, task-weighted selector infers hazards and ranks SLZs. Right: offline human-in-the-loop produces frozen rules for deployment.
• Figure 2: Illustration of the LLM–human-in-the-loop rule learning process.
• Figure 3: Qualitative outcomes in the Airsim. Two representative cases illustrate how the pipeline behaves under different scene conditions.
• Figure 4: Inference speed (FPS; higher is better). Boxes show the median and IQR; whiskers denote $1.5\times$IQR; dots are outliers. FPS is computed from per-trial latency ($\mathrm{FPS}=1000/\text{ms}$) under identical hardware and settings.
• Figure 5: Case study for a skateboard park. Green: feasible SLZ. Only the top-$k$ candidates ($k=10$) are shown in ours to avoid clutter; inference uses the full candidate set.