Impedance-Based Gaussian Processes for Modeling Human Motor Behavior in Physical and Non-Physical Interaction

Area of Science:

• Robotics
• Neuroscience
• Machine Learning

Background:

• Accurate modeling of human motor intention is crucial for predictive control in human-robot interaction.
• Current machine learning models often treat human motor behavior as a generic stochastic process, limiting prediction performance.
• Integrating prior knowledge about system structures can reveal unobservable intrinsic states for improved predictions.

Purpose of the Study:

• To present a novel method for modeling human motor behavior that incorporates neuroscientific insights.
• To improve prediction performance and generalization capabilities in human-robot interaction tasks.
• To leverage a priori knowledge of human arm impedance for enhanced interaction force prediction.

Main Methods:

• Developed a method modeling human motor behavior with a neuroscientifically supported impedance model of the human arm.
• Utilized a Bayesian framework with Gaussian Process (GP) priors for impedance elements and latent desired trajectories.
• Enabled regression of interaction forces by inferring a latent desired human trajectory, exploiting prior arm impedance knowledge.

Main Results:

• Validated the method using simulated data, demonstrating superior prediction performance compared to naive GP priors.
• Investigated the effects of GP prior parameterization and intention estimation.
• Evaluated generalization capabilities and the impact of training data sparsity through human participant experiments.

Conclusions:

• Derived correlations for an impedance-based GP model of human motor behavior that effectively utilizes prior knowledge.
• The model demonstrates robust prediction of interaction forces by inferring latent desired human trajectories.
• The approach shows effectiveness in both observed and unobserved regions of the input space, enhancing generalization.