Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hallucination-Aware Generative Pretrained Transformer for Cooperative Aerial Mobility Control

Hyojun Ahn, Seungcheol Oh, Gyu Seon Kim, Soyi Jung, Soohyun Park, Joongheon Kim

TL;DR

The paper tackles hallucination and safety challenges in transformer‑based UAV planning for last‑mile logistics by proposing SafeGPT, a two‑tier framework that couples a Global GPT for sector allocation with an On‑Device GPT for local routing, augmented by an RL safety filter and a dual replay buffer. It formalizes safety via a constrained MDP with a Lagrangian, an override mechanism, and a safety value function, ensuring battery conservation and route efficiency while mitigating unsafe plans. Empirical evaluation in a dynamic urban UAV scenario shows SafeGPT achieves 100% delivery success versus 98% for GPT‑Only, while reducing mean battery consumption from 76.4% to 61.1% and shortening travel distances, with a substantial decrease in hallucination‑driven actions. The work demonstrates that integrating semantic GPT reasoning with formal safety guarantees yields a practical, energy‑efficient solution for autonomous UAV logistics in uncertain environments.

Abstract

This paper proposes SafeGPT, a two-tiered framework that integrates generative pretrained transformers (GPTs) with reinforcement learning (RL) for efficient and reliable unmanned aerial vehicle (UAV) last-mile deliveries. In the proposed design, a Global GPT module assigns high-level tasks such as sector allocation, while an On-Device GPT manages real-time local route planning. An RL-based safety filter monitors each GPT decision and overrides unsafe actions that could lead to battery depletion or duplicate visits, effectively mitigating hallucinations. Furthermore, a dual replay buffer mechanism helps both the GPT modules and the RL agent refine their strategies over time. Simulation results demonstrate that SafeGPT achieves higher delivery success rates compared to a GPT-only baseline, while substantially reducing battery consumption and travel distance. These findings validate the efficacy of combining GPT-based semantic reasoning with formal safety guarantees, contributing a viable solution for robust and energy-efficient UAV logistics.

Hallucination-Aware Generative Pretrained Transformer for Cooperative Aerial Mobility Control

TL;DR

The paper tackles hallucination and safety challenges in transformer‑based UAV planning for last‑mile logistics by proposing SafeGPT, a two‑tier framework that couples a Global GPT for sector allocation with an On‑Device GPT for local routing, augmented by an RL safety filter and a dual replay buffer. It formalizes safety via a constrained MDP with a Lagrangian, an override mechanism, and a safety value function, ensuring battery conservation and route efficiency while mitigating unsafe plans. Empirical evaluation in a dynamic urban UAV scenario shows SafeGPT achieves 100% delivery success versus 98% for GPT‑Only, while reducing mean battery consumption from 76.4% to 61.1% and shortening travel distances, with a substantial decrease in hallucination‑driven actions. The work demonstrates that integrating semantic GPT reasoning with formal safety guarantees yields a practical, energy‑efficient solution for autonomous UAV logistics in uncertain environments.

Abstract

This paper proposes SafeGPT, a two-tiered framework that integrates generative pretrained transformers (GPTs) with reinforcement learning (RL) for efficient and reliable unmanned aerial vehicle (UAV) last-mile deliveries. In the proposed design, a Global GPT module assigns high-level tasks such as sector allocation, while an On-Device GPT manages real-time local route planning. An RL-based safety filter monitors each GPT decision and overrides unsafe actions that could lead to battery depletion or duplicate visits, effectively mitigating hallucinations. Furthermore, a dual replay buffer mechanism helps both the GPT modules and the RL agent refine their strategies over time. Simulation results demonstrate that SafeGPT achieves higher delivery success rates compared to a GPT-only baseline, while substantially reducing battery consumption and travel distance. These findings validate the efficacy of combining GPT-based semantic reasoning with formal safety guarantees, contributing a viable solution for robust and energy-efficient UAV logistics.

Paper Structure

This paper contains 17 sections, 9 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Modeling
  3. UAV Networks and Energy Modeling
  4. Generative Pretrained Transformer
  5. Reinforcement Learning with Safety Guarantees
  6. Algorithm Design
  7. GPT-Driven Planning and State Abstraction
  8. RL Constraints, Overriding, and Objective Function
  9. Dual Replay Mechanism and Integrated Algorithm
  10. Performance Evaluation
  11. Experimental Setup
  12. Episode Length and Customer Generation
  13. Comparison: GPT-Only vs. SafeGPT
  14. Simulation Procedure
  15. Performance Metrics
  16. ...and 2 more sections

Figures (5)

  • Figure 1: A conceptual overview of SafeGPT in a drone delivery scenario.
  • Figure 2: Two-tiered generative pretrained transformer architecture.
  • Figure 3: This prompt directs the Global GPT to generate high-level, safety-compliant decisions for drone sector allocation.
  • Figure 4: This prompt directs the On-Device GPT to produce control commands that are consistent with the prescribed safety constraints for the drone route plan.
  • Figure 5: Control performances of the UAV trained using the proposed algorithm and benchmark methods in dynamic environments.