Neural Network-based Control and Planning System Verification

Project Overview

This project focuses on the verification of neural network-based controllers or planners for safety-critical systems, particularly in the context of autonomous driving. The goal is to determine whether a neural network controller or planner can ensure that the system dynamics avoid unsafe regions while performing specific tasks, such as vehicle following. We aim to develop verifiable neural network controllers/planners and demonstrate their safety in these tasks, ensuring their reliability in real-world, high-stakes applications.

Objectives

  • Neural network controller design: develop neural network-based controllers for tasks in autonomous driving, with a focus on different scenarios.
  • Neural network planner design: design neural network-based planners for autonomous driving tasks, ensuring the system can navigate complex environments safely.
  • System dynamics modelling: define and model the system dynamics for the tasks, ensuring the system behavior is well-understood and controlled.
  • Safety analysis: conduct rigorous analysis and verification to ensure the neural network controller’s behavior is safe and robust across a range of operating conditions (e.g., varying traffic, road conditions, and sensor inputs).

Deliverables

  • Neural network controller: a trained neural network controller designed for vehicle-following tasks in autonomous driving systems.
  • Neural network planner: a neural network-based planner for autonomous driving tasks, ensuring safe and efficient navigation in complex environments.
  • Verification framework: a comprehensive framework for verifying the safety of neural network controllers/planners in dynamic systems.
  • Case studies: demonstrations of the verification techniques applied to real-world autonomous driving scenarios, showing the safety and effectiveness of the neural network controller.
  • Benchmarking: compare the performance of the neural network controller/planner for the formal verification community to evaluate the effectiveness of the proposed approach.