Verified Safe Neural Network Control for ESP32 in Closed-Loop Physical Systems
Objective: The goal of our project is to design and implement a verified neural network based controller running completely on an ESP32-S3 for a physical closed-loop system, here we are using a DC motor with encoder feedback. The thing we want to actually make is an on device runtime verification engine that uses interval arithmetic to mathematically guarantee that the controller’s output remains within the safe actuator bounds for all bounded sensor noise and disturbances. If the safety guarantee doesn’t work then the system automatically falls back to a provably safe PID controller. The entire motivation for this project revolves around the fact that modern control systems are increasingly using neural networks for their superior performance in complex, nonlinear environment, but NNs are notorious for producing unsafe outputs under adversarial inputs, sensor noise, or distribution shifts. The existing formal verification tools, for example: Reluplex, Marabou require high performance GPUs to run and cannot run on constrained environments such as a microcontroller like ESP32. We just want to merge the gap, if we can trust a neural network based controller on an embedded device when safety matters?