Mahya Mohammadi Kashani will give a talk about her work on Statistical Assessment of Plans via Probabilistic Optimization of Reliability for Underwater Robots.
SPEAKER: Mahya Mohammadi Kashani, PhD. fellow, ITU.
TITLE: Statistical Assessment of Plans via Probabilistic Optimization of Reliability for Underwater Robots
ABSTRACT: Autonomous Underwater Vehicles (AUVs) must operate reliably for long durations without human intervention, yet current task planning methods rely on abstract domain models that often lead to suboptimal real-world performance. This thesis introduces a novel probabilistic decision-making framework that enhances AUV planning through introspection (reasoning from the current state), retrospection (drawing on past experience), and prospection (forecasting future events). The framework is composed of three main components: Plan Generation, Plan Evaluation, and Plan Execution. Plans can originate from a knowledge base, traditional AI planners, or Large Language Models (LLMs).
First, a three-stage knowledge model is developed to improve fault recovery. It starts with a deterministic knowledge base that assesses the health of mission-critical components, then extends to a probabilistic layer estimating failure likelihoods. These are integrated into a ROS package that detects anomalies, diagnoses faults, proposes fixes, and selects the most likely successful recovery strategy. This is validated with both real AUVs and simulated ROVs encountering critical failures like thruster or camera malfunctions.
Second, a risk-aware planning approach is proposed to assess high-level plans using risk metrics—Variance, Entropy, Value at Risk (VaR), Conditional VaR (CVaR), and Entropic VaR (EVaR). Unlike traditional risk-neutral optimization, this method accounts for uncertainties in real-world execution. Simulations in tasks such as pipeline and subsea infrastructure inspection demonstrate its effectiveness, with Variance identifying the safest plan and EVaR/CVaR offering deeper insights into potential risks.
Finally, the thesis explores generating plans from past incident reports using LLMs. Incident Response Plans (IRPs) are converted into structured planning programs, producing datasets with high rates of parseability, solvability, and correctness. Plan diversity is evaluated using ML metrics like L2, Cosine, Wasserstein, and BERTScore. Fine-tuning CodeBERT for PDDL inputs improves embedding quality, while pre-trained Code-LLaMA performs well on L2 diversity. Despite some limitations in structural variation detection, this method shows promise in reducing the need for costly field trials.
Overall, this thesis advances AUV autonomy by integrating fault diagnosis, risk-aware planning, and LLM-based scenario generation to enhance mission reliability and planning robustness.