Abstract
We report the results obtained during the verification of Autosub6000, an autonomous underwater vehicle used for deep oceanic exploration. Our starting point is the Simulink/Matlab engineering model of the submarine, which is discretised by a compiler into a representation suitable for model checking. We assess the ability of the vehicle to function under degraded conditions by injecting faults automatically into the discretised model. The resulting system is analysed by means of the model checker MCMAS, and conclusions are drawn on the system's ability to withstand faults and to perform self-diagnosis and recovery. We present lessons learnt from this and suggest a general method for verifying autonomous vehicles.
Introduction: The Challenge of Deep-Sea Autonomy
Autonomous Underwater Vehicles (AUVs) like the Autosub6000 are high-stakes machines designed for deep oceanic exploration. Operating thousands of meters below the surface, often under ice, these vehicles are entirely out of communication range with their "mother ship." In such environments, a sensor failure or a low battery is a critical event; the vehicle must be intelligent enough to diagnose the problem and trigger a recovery or abort sequence autonomously.
While engineers traditionally use tools like Simulink for design, standard simulation cannot explore every possible "what if" scenario. This research bridges the gap between standard engineering models and formal verification—a mathematical approach that proves a system will always behave correctly under specified conditions. The focus is on verifying that a robot remains safe during faults (fault tolerance) and correctly identifies the cause of failures (self-diagnosability).
Methodology: The Integrated Verification Workflow
The researchers developed a systematic methodology that integrates industrial engineering tools with formal model checking. This workflow allows for the automated assessment of complex autonomous systems by bridging the gap between user-friendly design tools and rigorous analysis engines.
As shown in the workflow above, the process begins with existing Simulink/Stateflow models. These continuous models are translated into a discrete Multi-Agent System (MAS) formalism using a specialized compiler. This treats the AUV as a collection of independent "agents" (e.g., navigation, power, mission control) that communicate with one another.
A key innovation is the Integrated Fault Injection Approach. Rather than manually coding errors, a specialized compiler "mutates" the model by automatically injecting faults—such as a failing echosounder or low battery—based on specific timing rules. These mutated models are then analyzed using the MCMAS model checker against properties written in Temporal-Epistemic Logic (CTLK). This logic is particularly powerful because it can reason about "knowledge"—for instance, determining if the Mission Controller is actually aware that a specific sensor has failed.
Modeling and Fault Injection
The study utilized the formal model of the Autosub6000, which was derived from approximately 15,000 lines of Simulink code. The resulting formal model comprised 27 autonomous agents representing various subsystems, the environment, and human operators.
In collaboration with engineers, the authors selected 13 critical faults for injection, including PowerLow, AvoidTimerFail, and EchosounderDetects. This allowed the team to test the vehicle's logic against a wide array of failure scenarios that would be difficult to replicate in physical trials.
Key Insights and Results
The application of epistemic logic proved to be a significant advancement. While standard verification confirms if a state is reached, this method verifies if the robot's software is "aware" of the fault, a prerequisite for effective self-diagnosis. By treating the robot as a "society" of agents, the researchers could model how information and errors propagate through the system.
Performance and Safety Proofs
- Efficiency: The MCMAS tool verified 85 different safety and fault-tolerance specifications in under two hours on a standard desktop PC.
- Scalability: The methodology successfully managed a state space of over 47 billion reachable states, demonstrating that formal methods can handle real-world industrial complexity.
- Safety Verification: The system proved that the mission controller correctly enters a "termination exception" when power is low, provided the vehicle's status is consistent with operator observations.
- Bug Discovery: The process uncovered minor oversights in the original Autosub6000 design, which were subsequently corrected during the model refinement phase.
Conclusion and Future Directions
This work demonstrates that formal verification can be made accessible to engineers who are not logic experts. By "reversing" the compilation process, the researchers could show faulty behaviors discovered by the model checker directly within the original Stateflow diagrams.
However, challenges remain in the fidelity of abstraction. Discretizing continuous physics into discrete steps is complex and carries the risk of missing subtle timing issues. Future work aims to further automate complex diagnosability analyses and extend these methods to other autonomous domains, such as self-driving cars and aerial drones.
Summary updated