Technical Summary: ARIES-Mission2

Problem Statement

While Multimodal Large Language Models (MLLMs) demonstrate robust semantic understanding, their direct application to Unmanned Aerial Vehicle (UAV) mission generation faces critical limitations in spatial optimization and route planning. Current Vision-Language-Action (VLA) models often suffer from "semantic-to-physical" disconnects; they can extract target locations via zero-shot capabilities but lack the geometric reasoning required to generate efficient visiting sequences. This deficiency frequently results in redundant backtracking and physically infeasible navigation strategies. Fundamentally, multi-target UAV traversal constitutes a Traveling Salesperson Problem (TSP), an NP-Hard problem where standard MLLMs struggle to output combinatorially optimal solutions.

Methodology

The paper proposes ARIES-Mission2, a zero-shot Vision-Language-Action framework designed to decouple visual-semantic perception from physical route optimization. The system employs a "front-end inference + back-end optimization" architecture:

1. Front-End: Vision-Language Semantic Inference

The front end transforms unstructured natural language instructions and satellite imagery into precise geographic waypoints:

• Task Parsing: A large language model (DeepSeek-V3) parses natural language instructions to extract specific target categories (e.g., "buildings").
• Zero-Shot Grounding: A Vision-Language Model (Molmo-7B) locates these targets on the satellite image, outputting pixel coordinates.
• Geospatial Interpolation: A linear interpolation function maps 2D pixel coordinates to real-world GPS coordinates (latitude/longitude) using embedded image metadata, establishing a precise geographic baseline.

2. Back-End: Metaheuristic TSP Optimization

The back end addresses the combinatorial optimization of the visiting sequence using a TSP Solver:

• Cost Calculation: The system calculates physical flight distances between GPS waypoints using the Haversine formula to account for Earth's curvature.
• Algorithmic Competition: To overcome the limitations of raw VLM ordering, the framework evaluates four candidate routes in parallel:
  1. Raw VLM Order: The sequence directly generated by the vision-language model.
  2. PSO: Standard Particle Swarm Optimization.
  3. GPSO: An improved PSO introducing a dynamic inertia weight and a distance-based heuristic initialization to balance global exploration and local exploitation.
  4. IPSO: An improved PSO utilizing Tent Chaotic Maps for initialization (replacing pseudo-random numbers) and non-linear Lévy flight perturbations during updates to prevent premature convergence.
• Selection: The system selects the route with the minimum global distance among the four candidates.

3. Action Generation and Deployment

Unlike traditional VLA systems that rely on network-side LLMs to generate control code (risking latency and hallucinations), ARIES-Mission2 uses a local deterministic mission parser. It encapsulates the optimal GPS waypoints with preset flight parameters (e.g., altitude, landing commands) to generate standard MAVLink (QGC WPL 110) waypoint files within milliseconds for direct UAV execution.

Key Contributions

1. Decoupled Framework: The proposal of ARIES-Mission2, which separates VLM perception from metaheuristic TSP optimization to resolve inherent geometric reasoning defects in LLMs.
2. Algorithmic Innovation: The development of GPSO and IPSO variants that enhance global search capabilities and suppress premature convergence in high-dimensional flight tasks.
3. Efficiency Breakthrough: Demonstrating that coupling zero-shot visual grounding with lightweight route optimization significantly improves both flight efficiency and route quality compared to pure VLA baselines and human planning.

Experimental Results

Experiments were conducted on the UAV-VLPA-nano-30 benchmark, comprising 30 complex low-altitude flight tasks with diverse spatial topologies.

• Route Optimization:

  • Total Flight Distance: ARIES-Mission2 achieved a total distance of 62.43 km.
  • Comparative Performance: This represents a 21.6% reduction compared to the unoptimized VLA baseline (79.66 km) and a 9.5% improvement over manual human expert planning (69.00 km).
  • Per-Task Average: The optimized system averaged 2.08 km per task, outperforming the human average of 2.30 km.

• Computational Latency:

  • Total Workflow Time: The complete 30-task workflow took 575.40 seconds (approx. 9.5 minutes), averaging 19.18 seconds per task.
  • Speedup: This is approximately 3.6 times faster than the 35 minutes (2100 seconds) required for manual expert planning.
  • Component Breakdown: VLM inference dominates the runtime (19.02 s/task), while the TSP solver is highly efficient (0.16 s/task).

• Scalability:

  • Analysis of latency growth as the number of targets increased (from 4 to 20) showed that while VLM computation time grew steeply (approaching a 300% increase), the TSP solver's overhead remained stable (around 160%), confirming the scalability of the metaheuristic optimization module.

Significance and Claims

The paper claims that ARIES-Mission2 provides a novel paradigm for the efficient real-world implementation of multimodal large models in complex physical spaces. By decoupling semantic perception from physical optimization, the framework successfully bridges the gap between high-level semantic understanding and low-level geometric constraints.

The authors assert that this approach:

• Rectifies Trajectory Chaos: It effectively eliminates the redundant backtracking inherent in direct VLA inference.
• Surpasses Human Expertise: It achieves flight paths shorter than those planned by experienced human operators, demonstrating the potential of automated systems in spatial optimization.
• Ensures Safety and Efficiency: The local, deterministic generation of MAVLink missions ensures absolute physical safety and high computational efficiency, avoiding the latency and hallucination risks of cloud-based code generation.

The work concludes that this decoupled architecture is a viable solution for large-scale UAV mission generation, with future work planned to incorporate 3D terrain and dynamic obstacle avoidance mechanisms.