The University of Auckland

Project #38: Safe autonomy using synchronous AI and the F1/10 racing car

Back

Description:

 

Cyber-physical systems, such as autonomous vehicles, use interactive machine learning modules for decision making. Current approaches use a set of interacting neural networks, called neural network ensembles, to design such systems. These ensembles are realised using multithreading-based composition, which is not ideal for safety-critical systems.  Hence, for safety-critical systems such as autonomous vehicles, there is a need for rigorous approaches.

This project seeks to develop an alternative approach by reusing neural network ensembles, which are designed in the Keras library. After the application has been developed and the networks are trained, a compiler (recently developed by us), will be used to generate automatic C code, where the neural networks will operate synchronously. Our approach, thus, brings model driven engineering to the AI space. This will facilitate automatic static analysis to determine the worst case timing behaviour of the application, which is important for autonomy. 

The students will develop NN ensembles for autonomous vehicle application in Keras using deep neural networks. These will be then used to design an automated platooning system using the F1/10 racing vehicles (see: (1) http://www.madhurbehl.com/f110.html (2) https://www.youtube.com/watch?v=vgEyvazwrU8)

 

References

Roop PS, Pearce H, Monadjem K. Synchronous neural networks for cyber-physical systems. In2018 16th ACM/IEEE International Conference on Formal Methods and Models for System Design (MEMOCODE) 2018 Oct 15 (pp. 1-10). IEEE.

Allen N, Raje Y, Ro JW, Roop P. A compositional approach for real-time machine learning. InProceedings of the 17th ACM-IEEE International Conference on Formal Methods and Models for System Design 2019 Oct 9 (pp. 1-5).

Yang X, Roop PS, Pearce H, Woo JR, A COMPOSITIONAL APPROACH USING KERAS FOR NEURAL NETWORKS IN REAL-TIME SYSTEMS, Design Automation and Test in Europe (DATE), March 2020, Grenoble, France.

R. Ivanov, J. Weimer, R. Alur, G.J. Pappas, and I. Lee. Case Study: Verifying the Safety of an Autonomous Racing Car with a Neural Network Controller, arXiv, 2019.

R. Ivanov, J. Weimer, R. Alur, G.J. Pappas, and I. Lee. Verisig: Verifying safety properties of hybrid systems with neural network controllers, 22nd ACM International Conference on Hybrid Systems: Computation and Control, 2019.

 

Keras https://keras.io/ 

 

Type:

Undergraduate

Outcome:

  1. Design and verification of safe neural networks for autonomous platooning systems
  2. Design of a autonomous vehicle platoon using the F1/10 racing cars

Prerequisites

None

Specialisations

Categories

Supervisor

Co-supervisor

Team

Lab

Lab allocations have not been finalised