LASSA Architecture-Based Autonomous Fault-Tolerant Control of Unmanned Underwater Vehicles

Researchers propose LASSA, an LLM-based autonomous control architecture for unmanned underwater vehicles that combines large language models with physical constraint verification to enable fault-tolerant operation in communication-limited environments. Lake experiments demonstrate the system successfully detects faults, replans missions, and maintains operational safety without false alarms.

This research addresses a critical gap in autonomous systems design: deploying large language models in safety-critical environments where hallucinations could cause operational failures. The LASSA architecture represents a pragmatic engineering solution that doesn't attempt to eliminate LLM unreliability but rather contains it through layered verification mechanisms. The dual closed-loop control system—separating high-level decision-making from real-time control execution—reflects mature systems engineering principles adapted for AI integration.

The significance lies in demonstrating that LLMs can provide genuine value in autonomous vehicles beyond simple language tasks. By leveraging their reasoning capabilities for fault diagnosis and mission replanning while constraining outputs through a solver that verifies physical feasibility, the system achieves interpretability without sacrificing adaptability. This contrasts with purely rule-based legacy systems that struggle with unforeseen failures.

For the broader autonomous systems industry, this work validates a promising integration pattern: LLMs as strategic decision-makers with engineering guardrails rather than end-to-end controllers. The lake trials showing successful fault detection, appropriate mission parameter adjustment (4m to 12m turning radius, 2kn to 1kn speed reduction), and constraint satisfaction on first invocation indicate practical viability. Zero false alarms during normal operation suggest the architecture avoids introducing new failure modes.

The framework has implications for underwater robotics, offshore operations, and any autonomous domain requiring persistent operation in degraded communications. Developers monitoring this space should track whether similar architectures emerge in other safety-critical domains like aerospace or industrial automation.

• →LASSA architecture constrains LLM outputs through physical feasibility verification to prevent hallucinations in autonomous underwater vehicles
• →Dual closed-loop design separates strategic decision-making from real-time control, balancing intelligence with operational safety
• →Lake trials demonstrate successful fault detection, autonomous mission replanning, and zero false alarms in normal conditions
• →The system adjusts operational parameters in response to detected faults without human intervention or hard-coded rules
• →Architecture pattern may become foundational for deploying LLMs in other safety-critical autonomous systems