Generating Out-Of-Distribution Scenarios Using Large Language Models

This project explores the use of Large Language Models (LLMs) to tackle the critical challenge of addressing Out-Of-Distribution (OOD) scenarios in autonomous driving, which are essential for ensuring safety and reliability under unpredictable conditions. Leveraging the zero-shot generalization and reasoning capabilities of LLMs, we introduce a framework that generates diverse OOD scenarios as a branching tree, with each branch representing a unique case. These scenarios are simulated in the CARLA simulator through automated scene augmentation aligned with textual descriptions. We evaluate the framework using a diversity metric and a novel "OOD-ness" metric to quantify deviation from typical urban conditions. Additionally, we assess the ability of Vision-Language Models (VLMs) to interpret and navigate these scenarios. This work highlights the transformative potential of language models in enhancing the safety validation of autonomous vehicles.

Autonomous Driving in Urban Environments Under Uncertain Conditions

This project introduces a two-level hierarchical architecture for controlling autonomous vehicles in complex urban environments, prioritizing collision avoidance, traffic rule adherence, and real-time performance. By integrating Signal Temporal Logic (STL) into Model Predictive Control (MPC) at the top level and leveraging detailed feedback from the bottom level, the system bridges the gap between simplified control models and real-world dynamics. This closed-loop approach enhances safety and reliability, addressing a critical challenge in autonomous driving. Simulations in the CARLA simulator demonstrate the method's effectiveness and efficiency compared to existing solutions, highlighting its potential to advance urban autonomous vehicle navigation. By tackling the discrepancies between simplified models and real-world conditions, this framework provides a robust foundation for safer and more dependable autonomous systems.

Interpretable Classification of Time-Series Data Using Decision Trees

Classifying time-series data is critical for the analysis and control of autonomous systems, including robots and self-driving cars, where interpretability is essential for safety and trust. Existing temporal logic-based learning methods often fall short in real-world applications due to inaccuracies or the complexity of the generated formulae. To address these challenges, we present Boosted Concise Decision Trees (BCDTs), a novel approach that combines an ensemble of simplified decision trees to produce Signal Temporal Logic (STL) classifiers. This method enhances classification accuracy while prioritizing interpretability by generating concise and comprehensible formulae. The effectiveness of BCDTs is demonstrated through naval surveillance and urban-driving case studies, highlighting its practical significance and potential for real-world impact.

Deep Reinforcement Learning in Complex Evnironments with Interpretable Specifications

This project addresses the challenges of robot navigation in cluttered environments with unknown dynamics by leveraging deep reinforcement learning (DRL) to handle tasks specified through Linear Temporal Logic (LTL) formulas. Traditional DRL approaches struggle with exploration due to sparse rewards, a common issue in obstacle-dense settings. To tackle this, we propose a novel framework that integrates path planning-guided reward schemes with sampling-based methods to enhance exploration and ensure efficient task completion. By decomposing complex LTL missions into distributed sub-goals, our approach significantly improves the performance and adaptability of robots in achieving intricate missions, demonstrating the transformative potential of reinforcement learning in robotics.