Confidence-Infused Operator Behavioral Intent Inference for Teleoperated Robots with Uncertainty in Goals

This paper introduces a Bayesian intent inference method that utilizes indirect signals from the control interface to infer the operator's intended navigation goals. Aimed at minimizing intent prediction errors and addressing unknown or changing intents, the method incorporates a dynamic operator behavior model with confidence-infused Bayesian evidence. This model employs real-time probabilistic predictions of human inputs, integrates the robot's confidence in its cost function features, and adjusts these features as required. Validation using data from two real-world scenarios shows that our method outperforms state-of-the-art approaches in reducing prediction error rates. Notably, it significantly improves inference accuracy under partial intent knowledge while maintaining low incorrect inference rates.