Automating the Extraction or Learning of Hybrid Automata from Simulation Data

Project Overview

The goal of this project is to explore how to automatically extract or learn hybrid automata from simulation data related to autonomous driving systems (e.g., ACC). The project can involve designing a tool or proposing an algorithm or approach, leveraging both active learning (query-based learning) and passive learning (data-driven learning) techniques. These learned automata models will be tested for safety, efficiency, and robustness in new driving scenarios to ensure their effectiveness in real-world environments.

Objectives

  • Data collection: simulate various driving scenarios related to autonomous driving systems (e.g., ACC) using simulation platforms (e.g., Simulink), and collect sensor data, vehicle states, and control commands.
  • Learning algorithm application: design and implement an algorithm that automatically extracts and translates models from simulation platforms (such as Simulink) into hybrid automata, capturing the state transitions and decision-making logic of the system.
  • Safety and efficiency evaluation: test the learned automata models in new driving scenarios, evaluating their safety (collision rate, traffic law compliance), efficiency, and robustness.
  • Real-world validation: validate the performance of the learned models in real-world simulations and further optimize them.
  • Evaluation metrics: accuracy of the translation from autonomous driving system behavior to hybrid automata, including the correctness of state transitions and decision-making logic.

Deliverables

  • Automaton model: a learned hybrid automaton representing the decision-making process of the autonomous driving system.
  • Learning algorithms: active learning and passive learning algorithms for automatically extracting hybrid automata from simulation data.
  • Learning pipeline: a fully functional pipeline for collecting simulation data and learning hybrid automata models.
  • Evaluation report: a detailed report evaluating the safety, efficiency, and robustness of the learned models.
  • Codebase: complete code and installation instructions for the learning framework.