Mutual Adaptation and Influence: Review of Latent Dynamics Models in Human–Robot Interaction

Recent advances in robotics have enabled more dynamic and interactive roles when working with humans. This progress is enabled in part by latent state models, which allow the robot to predict human actions in a compact and tractable way. This paradigm has allowed robots not only to respond and adapt to human behavior but also to anticipate and plan proactively through the use of latent dynamics models. This review explores the use of such models in the field of human–robot interaction. Existing literature reveals three classes of latent dynamics models: Bayesian, Markovian, and encoded. To connect these works, we synthesize a unified framework consisting of a prediction and control step. This framework shows how the traditional one‐way adaptation extends to mutually adaptive behaviors by using latent dynamics in the prediction step. Similarly, we show how influence is an extension of mutual adaptation by using latent dynamics in the control step. We then review state‐of‐the‐art approaches to mutual adaptation and influence for the three latent dynamics classes and discuss emergent properties as they relate to application, update frequency, and specific considerations for mutual adaptation or influence. Using this review, we highlight gaps in the current literature and propose future directions.