About

Neural networks (NNs) are useful for many controls problems. However, while they may behave well for a nominal input, NNs are not generally robust to perturbations to that input, which may exist due to noise or adversarial attacks. For this reason, as NNs are applied to safety-critical systems such as mobile robotics, it is necessary to develop algorithms that can provide safety assurances about these systems. This project developed a backward reachability analysis approach that can determine what states a system would have to come from to end up in a given target set. Our approach can be used to guarantee safety for the system by setting the target set as an obstacle and ensuring that the system is not in the backprojection set, i.e., the set of states that will collide with the obstacle.

Software

https://github.com/neu-autonomy/nfl_veripy

Sponsor

This research is funded by Ford Motor Company.

Related Publications

• Rober, N., Everett, M., and How, J. P., “Backward reachability analysis for neural feedback loops,” 2022 IEEE 61st Conference on Decision and Control (CDC), IEEE, 2022, pp. 2897–2904.
• Rober, N., Katz, S. M., Sidrane, C., Yel, E., Everett, M., Kochenderfer, M. J., and How, J. P., “Backward reachability analysis of neural feedback loops: Techniques for linear and nonlinear systems,” IEEE Open Journal of Control Systems, 2023.
• Rober, N., Everett, M., Zhang, S., and How, J. P., “A hybrid partitioning strategy for backward reachability of neural feedback loops,” 2023 American Control Conference (ACC), IEEE, 2023, pp. 3523–3528.