Simultaneous State Estimation and Online Model Learning in a Soft Robotic System

Researchers have developed a new AI system that allows soft robots to simultaneously estimate their current state and learn a model of their own bending stiffness online, using only base force measurements and a nominal model. The method combines a marginalized particle filter with a Gaussian Process model, enabling real-time learning and improved prediction accuracy. Tests on real-world soft-robot data show it reduces multi-step forward-prediction errors while accurately estimating the robot's pose.