Technical Summary: Complex Autonomous UAV Task Execution and Decision-Making With s(CASP)

Problem Statement

The deployment of autonomous Unmanned Aerial Vehicles (UAVs) in dynamic, real-world environments faces significant challenges regarding safety, explainability, and adaptability. While deep learning approaches have demonstrated strong empirical performance in perception and navigation, they function as "black boxes" that are difficult to explain, verify, or modify once deployed. These models often require extensive offline training and lack the ability to provide justifiable reasoning for their decisions, which undermines trust in safety-critical missions such as disaster response, inspection, and agriculture. Furthermore, traditional logic-based approaches often suffer from combinatorial explosions during the "grounding" process, making them computationally infeasible for complex, real-time reasoning tasks.

Methodology

The authors propose and fully implement a symbolic state-centered UAV agent architecture known as VECSR-A (Virtually-Embodied Common Sense Reasoner for Autonomy). This system integrates s(CASP), a goal-directed, constraint-based Answer Set Programming (ASP) system, with a high-fidelity simulation environment.

Core Components

1. s(CASP) Reasoning Engine: Unlike traditional ASP solvers that require full grounding of the entire program, s(CASP) performs top-down, goal-directed search. It computes partial answer sets to satisfy specific queries, enabling the handling of complex reasoning problems without combinatorial explosion. It supports:

   • Nonmonotonic reasoning: Using default negation (negation-as-failure) and strong negation.
   • Abducibles: Allowing the system to assume facts (e.g., a location is safe) in the absence of contradictory evidence.
   • Constraints: Pruning answer sets that violate safety rules (e.g., flying into a no-fly zone).
   • Explainability: Generating proof trees that serve as verifiable justifications for every action.

2. VECSR-A Architecture: The system acts as an orchestrator that bridges the simulated environment and the symbolic reasoning engine.

   • Perception-to-Logic: Machine learning models or ground-truth data from the simulation (e.g., camera feeds) are translated into logic predicates representing the current state.
   • Task Decomposition: High-level goals (e.g., "deliver a package") are hierarchically decomposed into low-level primitive actions (e.g., "move," "pick up," "drop off").
   • Execution Loop: The agent queries the s(CASP) server with the current state and goal. The server returns a set of legal, suggested actions with justifications. The agent executes the chosen action in the simulation, updates the state, and repeats the cycle.

3. Simulation Environment: The framework is evaluated using Project AirSim running on Unreal Engine 5. This platform provides a realistic, high-fidelity city environment with dynamic objects and complex spatial constraints, ensuring deterministic reproducibility for evaluation.

Key Contributions

The paper presents three primary contributions:

1. Full Implementation of VECSR-A: The authors fully realized the VECSR-A architecture to support multi-step autonomous task execution via s(CASP) goal-directed reasoning within the high-fidelity Project AirSim environment.
2. Demonstration of Symbolic Advantages: The work demonstrates the benefits of symbolic commonsense reasoning and task decomposition over deep learning models for drone control, specifically regarding explainability and the ability to adapt to new constraints without retraining.
3. Empirical Validation: The framework is validated through runtime analysis across five distinct mission scenarios, confirming the system's ability to perform real-time decision-making in dynamic urban environments.

Experimental Results

The system was evaluated on five core tasks: Object Inspection, Debris Detection, Trespassers, Spraying, and Transport/Delivery.

• Performance: The system successfully completed all designated tasks, navigating the map, identifying objects, and adhering to complex constraints (e.g., avoiding spraying garden beds if a person is nearby or if the bed is obstructed).
• Runtime Metrics:
  • The s(CASP) reasoning engine generates actionable plans in approximately 1.2 to 1.3 seconds per cycle.
  • Total task completion times ranged from roughly 22 to 39 seconds, with the majority of time spent on physical drone movement rather than computational overhead.
• Ablation Studies:
  • Navigation: The system successfully navigated to target locations in all test scenarios, handling collisions effectively.
  • Interaction: The "suggested action" component correctly identified necessary environment interactions in all scenarios.
  • Constraints: The "legal action" component correctly identified the legality of actions in all scenarios, ensuring compliance with spatial and social constraints.

Significance and Claims

The paper claims that the VECSR-A framework provides a flexible, scalable approach to explainable autonomy. By leveraging s(CASP), the system guarantees that decisions are correct, justifiable, and based on commonsense reasoning.

• Explainability: Unlike deep learning models, VECSR-A produces explicit logical proof trees for every action, allowing human operators to understand exactly why a decision was made or skipped. This is critical for mission-critical operations.
• Adaptability: The system can adapt to new mission parameters or constraints (e.g., "do not spray if a person is present") simply by modifying rules in the knowledge base, without the need for costly retraining or architectural changes.
• Real-Time Viability: The goal-directed nature of s(CASP) allows for near real-time decision-making, making it suitable for deliberative tasks in dynamic environments where the state changes between decisions.

The authors conclude that while the current one-second reasoning cycle is well-suited for deliberative tasks, future work may need to address millisecond-level reactive scenarios and integrate probabilistic models to handle sensor noise and uncertainty. Nevertheless, the framework establishes a robust foundation for autonomous systems where reliability and explainability are paramount.