Researcher Tests If Robot Actions Can Be Predicted From Microcontroller Timing Alone

An independent researcher is investigating whether timing side-channel attacks can reveal the actions of multi-agent reinforcement learning (MARL) policies deployed on edge hardware. The study focuses on an ESP32-S3 microcontroller running small neural network policies, where an attacker measures only inference duration and network packet timing without access to inputs, weights, or activations. Experiments span three environments: a custom cooperative grid navigation task, the classic CartPole benchmark, and the PettingZoo MPE Simple Spread scenario. Policies trained via PPO are exported to TFLite and flashed onto the microcontroller, with a custom pipeline used to collect and quantify timing leakage. The research highlights a potential security risk for edge-deployed autonomous systems such as drones, warehouse robots, and IoT networks, where physical proximity could allow adversaries to exploit such timing signals.