💻 computer science

NEUROSYMLAND: Neuro-Symbolic Landing-Site Assessment for Robust and Edge-Deployable UAV Autonomy

The paper presents NEUROSYMLAND, a neuro-symbolic framework that combines lightweight visual perception with explicit symbolic reasoning to achieve robust, interpretable, and edge-deployable landing-site assessment for autonomous UAVs, demonstrating superior performance and bounded resource usage across extensive simulated and hardware-in-the-loop evaluations.

Original authors: Weixian Qian, Tianyi Yang, Sebastian Schroder, Yao Deng, Jiaohong Yao, Xiao Cheng, Richard Han, Xi Zheng

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Weixian Qian, Tianyi Yang, Sebastian Schroder, Yao Deng, Jiaohong Yao, Xiao Cheng, Richard Han, Xi Zheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a drone that needs to land in a messy, unpredictable backyard. It can't just "look" at the ground and guess; it needs to be absolutely sure the spot is safe, flat, and free of people or pets.

The paper introduces NEUROSYMLAND, a new system designed to help drones make these landing decisions safely and transparently, even on small, low-power computers (like the ones inside the drone itself).

Here is how it works, broken down into simple concepts and analogies:

1. The Problem: The "Black Box" Dilemma

Most current drones use "deep learning" (AI that learns by looking at thousands of pictures). Think of this like a student who memorized the answers to a test but doesn't understand the math.

The Issue: If the student sees a slightly different question (a weird terrain or lighting), they might guess wrong. Worse, if they get it wrong, you can't ask them why. They just give an answer.
The Risk: In a real-world emergency, if a drone lands in a puddle or hits a person, we need to know exactly why it made that mistake so we can fix it.

2. The Solution: The "Architect and the Lawyer"

NEUROSYMLAND splits the job into two distinct roles, like a construction team:

Role A: The Architect (The "Neuro" Part)
This is the drone's eyes. It looks at the camera feed and draws a rough map. It identifies things like "grass," "rocks," "people," and "flat pavement."
- Crucial Detail: Instead of just saying "Landing Score: 8/10," it builds a Probabilistic Semantic Scene Graph (PSSG). Think of this as a structured checklist or a digital filing cabinet. It doesn't just see a blob of green; it files a note: "There is a patch of grass here, it is 2 meters wide, and it is next to a tree." It also notes how confident it is about each item.
Role B: The Lawyer (The "Symbolic" Part)
This is the decision-maker. It doesn't "guess." It follows a strict, written rulebook (like a legal contract).
- The Rulebook: "IF there is a person nearby, THEN do not land." "IF the ground is not flat, THEN do not land." "IF the area is too small, THEN do not land."
- The Magic: These rules are written in plain logic, not hidden inside a neural network. If the drone lands somewhere, the "Lawyer" can show you the exact page of the rulebook it used to make that decision.

3. How They Work Together

The Architect looks at the world and fills out the checklist (the PSSG).
The Lawyer reads the checklist. It checks the facts against the rulebook.
The Decision: The Lawyer says, "Okay, this spot has grass (Fact), but it is next to a dog (Fact). The rule says 'No landing near dogs.' Therefore, this spot is unsafe."

4. The "Human-in-the-Loop" Refinement

The paper mentions a clever way to write the rules. They used an AI (a Large Language Model) to help draft the initial rules, but only before the drone ever flies.

Imagine a human expert and an AI assistant sitting down to write the rulebook together. The AI suggests, "Maybe we should add a rule about dogs?" The human expert checks it, says "Yes, but make it stricter," and writes the final rule.
Once the drone is flying, the AI assistant is gone. The drone runs on the fixed, human-approved rulebook. This ensures the drone is fast, predictable, and doesn't get confused by the AI "hallucinating" new rules mid-flight.

5. Why This is Better (The Results)

The researchers tested this system in 72 different simulated landing scenarios (like a video game test drive) and on real hardware (a small computer called a Jetson Orin Nano).

Success Rate: NEUROSYMLAND successfully identified safe landing spots 61 out of 72 times. This was better than other top systems, which only succeeded 37 to 57 times.
Transparency: If the drone fails, you can look at the "Lawyer's" notes and see exactly which rule was broken.
Speed & Efficiency: Even though it does two steps (Architect + Lawyer), it runs fast enough for a drone. The "Lawyer" part is so fast it takes almost no time (less than 2% of the total processing time). The heavy lifting is done by the "Architect" (the camera), which is necessary anyway.
Adaptability: If you want the drone to act differently (e.g., "Rescue Mission" vs. "Delivery Mission"), you just change the weights in the rulebook. You don't have to retrain the whole AI brain.

Summary

NEUROSYMLAND is like giving a drone a pair of eyes (to see the world) and a strict, written rulebook (to make decisions).

It avoids the "black box" problem where AI guesses without explanation.
It is robust enough to handle messy, real-world terrain better than previous methods.
It is light enough to run on a small drone computer without needing a supercomputer.

The paper concludes that this "Neuro-Symbolic" approach is a major step toward making autonomous drones safe enough to fly near people and infrastructure, because their decisions are logical, explainable, and reliable.

Technical Summary: NEUROSYMLAND

Problem Statement

Safe landing-site assessment (LSA) in unstructured environments remains a critical bottleneck for autonomous UAV deployment. Existing marker-free approaches predominantly rely on end-to-end learning models that regress landing feasibility directly from geometric and semantic features. While effective at identifying obvious hazards, these methods suffer from three primary limitations:

Lack of Transparency: They provide limited interpretability regarding safety decisions, making systematic debugging and certification difficult.
Rigidity: Updating operational constraints (e.g., mission-specific policies or safety margins) often requires retraining or heuristic redesign.
Edge Constraints: Recent LLM/VLM-based methods offer higher-level reasoning but typically depend on black-box inference or heavy foundation models that exceed the memory and compute budgets of edge-deployed robotic systems.

Methodology

The paper proposes NEUROSYMLAND, a neuro-symbolic framework that explicitly decouples probabilistic world modeling from fully symbolic mission-level reasoning. The system operates in two distinct phases:

1. Online Perception and World Modeling

Input Processing: A monocular image is processed by a lightweight, INT8-quantized SegFormer-B0 segmentation model with geometric post-processing.
Probabilistic Semantic Scene Graph (PSSG): Instead of outputting a raw landing mask, the system constructs a PSSG ( $G = (V, E)$ $G = (V, E)$ ).
- Nodes: Represent segmented regions with attributes (semantic confidence, area, centroid, shape descriptors, flatness proxies).
- Edges: Encode spatial relations (e.g., adjacent to, near to, contain) computed deterministically from geometry.
- Uncertainty: The graph serves as an explicit, inspectable world model where node attributes store semantic posteriors and edges encode relation confidences.

2. Symbolic Reasoning and Decision Making

Rule Engine: The decision layer utilizes SCALLOP, a probabilistic logic programming engine. It executes a fixed, pre-compiled set of symbolic rules over the PSSG.
Offline Rule Construction: Rules are synthesized offline using an LLM (OpenAI o3) as an assistant to translate expert constraints and failure-case feedback into candidate SCALLOP rules. These are iteratively refined via human-in-the-loop (HITL) validation but no LLM inference occurs at runtime.
Inference Process:
- The engine infers hazard(x) and safe(x) for candidate regions based on predicates like is flat surface(x) and near to(x, person).
- Multi-frame Validation: Candidates are validated over a temporal window ( $T$ ) to ensure geometric stability and absence of transient hazards.
- Mission-Specific Ranking: A utility function ranks admissible regions based on mission weights (e.g., proximity to target vs. safety margins), producing a final landing proposal.

Key Contributions

Modular Neuro-Symbolic Architecture: The framework separates learned perception from symbolic reasoning, enabling structured decision control under uncertainty while maintaining transparency.
Deterministic Runtime Reasoning: By executing compiled SCALLOP rules over a PSSG, the system produces traceable, debuggable, and reconfigurable decisions without neural inference in the decision layer.
LLM-Assisted Offline Engineering: The paper introduces a workflow where LLMs assist only in drafting rules offline. The final runtime system is a fixed, expert-validated rule set, ensuring deterministic behavior suitable for certification.
Edge-Feasible Evaluation: The system is evaluated on embedded hardware (Jetson Orin Nano), demonstrating that explicit reasoning adds negligible latency compared to perception costs.

Experimental Results

The authors evaluated NEUROSYMLAND across 72 simulated landing scenarios (Software-in-the-Loop) and 100 hardware-in-the-loop (HIL) trials.

Effectiveness (SIL)

Success Rate: NEUROSYMLAND achieved 61/72 successful assessments, outperforming competitive baselines: PEACE (57/72), SafeUAV (47/72), and LLMExplain (37/72).
Robustness: The system demonstrated superior robustness to fragmented semantics and spurious labels by reasoning over region-level evidence rather than pixel-level masks.
Ablation: Replacing probabilistic rules with deterministic First-Order Logic (NEUROSYMLAND-DetFOL) resulted in 58/71 successes, indicating that the probabilistic grounding of the PSSG further enhances decision quality.

Efficiency (HIL)

Latency: The system achieved an end-to-end latency of 1,043.3 ms (~0.96 FPS) on a Jetson Orin Nano.
Computational Breakdown:
- Perception and PSSG construction accounted for ~94% of the total latency.
- Symbolic reasoning contributed only ~2% (20.0 ms), confirming that the reasoning overhead is negligible.
Resource Usage: The system operated stably within embedded constraints, utilizing 76.9% CPU, 52% GPU, and 5.6 GB memory, with a power consumption of 12.3 W.

Interpretability

Qualitative analysis showed that NEUROSYMLAND provides explicit provenance traces, attributing decisions to specific rules (e.g., near person, near obstacle). Unlike baselines that output dense masks or single points, NEUROSYMLAND allows for mission-adaptive reconfiguration (e.g., switching between "Safe," "Emergency," and "Rescue" modes) simply by adjusting ranking weights without retraining.

Significance and Claims

The paper claims that NEUROSYMLAND demonstrates the feasibility of deploying robust, interpretable landing-site assessment stacks on edge-constrained UAV hardware. Its primary significance lies in:

Decoupling Perception and Control: It allows for mission reconfiguration and systematic debugging by modifying symbolic rules rather than retraining neural networks.
Bounded Edge Resource Usage: It proves that explicit symbolic reasoning does not impose a prohibitive computational burden on embedded systems, as the cost is dominated by perception.
Transparency: The system offers human-readable reasoning paths and traceable safety scores, addressing the "black-box" nature of purely learning-based approaches.

The authors acknowledge that the system's performance is fundamentally bounded by the quality of the underlying perception (segmentation errors propagate to the PSSG) and that current scenarios lack significant dynamic elements. Future work is directed toward improving world-model reliability, extending to dynamic environments, and integrating the assessment module with full flight-control stacks for closed-loop trials.