Deep Active Inference in Physical Human-Robot Interaction: Balancing Exploration and Goal-Directed Behavior

Jefimija Borojevic, Gabriel W. Haddon-Hill, Juan Sandoval, Shingo Murata

Abstract: In physical Human-Robot Interaction (pHRI), a significant challenge lies in the unpredictable nature of human behavior, which can introduce a high level of uncertainty during an interaction with a robot. While traditional force-based control laws lack high-level reasoning, learning-based methods do not treat perception and action uniformly to decrease uncertainties about human intention. This paper presents a pHRI framework based on Active Inference (AIF) for planning and decision-making, which guides a robot to balance goal-directed behavior and the exploration of the human intention. The framework integrates a 1D-CNN based Conditional-Variational Autoencoder (CVAE) architecture for Expected free energy (EFE) computation and policy selection with a Cartesian impedance controller to allow the robot to adapt its motion according to the selected policy. Additionally, this paper investigates the influence of preference precision on the robot’s behavior. The pHRI experiment includes pushing, pulling and no-applied interaction scenarios. The results show that the robot naturally favors goal-directed, high-stiffness behavior when it is undisturbed, while it prefers exploration behavior for lower preference precision values and goal-directed behavior for higher preference precision values during pushing and pulling interaction scenarios. The findings of this research demonstrate that the preference precision parameter significantly influences the process of minimizing uncertainty of human behavior, enabling the robot to adaptively balance exploration and goal-directed behavior in pHRI scenarios.