Learning and Blending Robot Hugging Behaviors in Time and Space

TL;DR

The paper tackles the challenge of making robot hugging interactions adaptive to human partners by blending multiple sub-actions in real time. It introduces Blending Bayesian Interaction Primitives (B-BIP), a probabilistic framework that extends Bayesian Interaction Primitives (BIP) to handle multiple interaction primitives and arbitrary switching, using phase-aware latent representations and ensemble Kalman filtering. Through a large dataset of hugging demonstrations and a live participant study, B-BIP demonstrates lower prediction error and more favorable human responses than baselines (BIP, ProMP, LSTM), especially during switching hugs. This work lays a foundation for natural, responsive social robots capable of real-time blending of interaction modes, with potential impact across human-robot interaction domains.

Abstract

We introduce an imitation learning-based physical human-robot interaction algorithm capable of predicting appropriate robot responses in complex interactions involving a superposition of multiple interactions. Our proposed algorithm, Blending Bayesian Interaction Primitives (B-BIP) allows us to achieve responsive interactions in complex hugging scenarios, capable of reciprocating and adapting to a hugs motion and timing. We show that this algorithm is a generalization of prior work, for which the original formulation reduces to the particular case of a single interaction, and evaluate our method through both an extensive user study and empirical experiments. Our algorithm yields significantly better quantitative prediction error and more-favorable participant responses with respect to accuracy, responsiveness, and timing, when compared to existing state-of-the-art methods.

Figures (6)

• Figure 1: Robot dressed as a plush bear hugs a human partner.
• Figure 2: An overview of B-BIP. Top: training demonstrations (left) are decomposed into a latent space (middle) and transformed into an ensemble of samples (right). Bottom: observations are collected during a live interaction (left) which is used to perform filtering with the learned ensemble (middle) and produce a response trajectory (right).
• Figure 3: An example of a left-high to right-high interaction. Left: The participant starts with a left-high interaction. Middle: When switching to the right-high hug, the robot responds accordingly. Right: The participant hugs the robot.
• Figure 4: Top left: The observed $z$ positions of the participant's hands and the robot's end effectors during a right-high to left-high interaction. Bottom left: The corresponding interaction class weights for each hug type, with vertical area equal to the class probability. Right: The trajectory of the observed DoF (black arrows) projected to the reduced-rank LDA space, overlaid on the distributions (circles) for each hug type.
• Figure 5: Distribution of scores for the three questions- which are used for hypothesis tests- after switching interactions.