Fault Detection and Response for Safe Control of Artificial Muscles in Soft Robots

Abstract

Robots built from soft materials have the potential for intuitively-safer interactions with humans and the environment. However, soft robots’ embodiments have many sources of failure that could lead to unsafe conditions in closed-loop control, such as degradation of sensors or fracture of actuators. This letter proposes a fault detection system for sensors attached to artificial muscle actuators that satisfies a formal safety condition. Our approach combines redundant sensing, model-based state estimation, and Gaussian process regression to determine when one sensor’s reading statistically diverges from another, indicating a fault condition. We apply the approach to electrothermal shape memory alloy (SMA) artificial muscles, demonstrating that our method prevents the overheating and fire damage risk that could otherwise occur. Experiments show that when the muscle’s nominal sensor (temperature via a thermocouple) is fractured from the robot, the redundant sensor (electrical resistance) combined with our method prevents violation of state constraints. Deploying this system in real-world human-robot interaction could help make soft robots more robust and reliable.