We released Predictive-Coding-Based Autonomous Regulation of Internally Generated and Externally Coupled Processing in Human-Robot Interaction on arXiv. The paper is first-authored by Henrique Oyama and co-authored with Prof. Jun Tani. It has been submitted to IEEE Transactions on Cognitive and Developmental Systems.
A robot that physically interacts with a person must keep deciding how much to rely on its learned internal predictions and how much on the sensory input it is actually receiving. Relying too much on predictions prevents it from accommodating the partner; relying too much on sensory input erodes the structure of its learned behavior.
This work focuses on the meta-prior of PV-RNN, a predictive-coding variational RNN, which controls how strongly posterior inference is constrained by learned prior dynamics. We introduce an online mechanism that switches among predefined meta-prior settings according to the history of reconstruction error accumulated over recent interaction, letting the robot itself regulate its reliance on internal predictions versus external sensory evidence.
Across three physical human-robot interaction tasks (fixed structured, changing structured, and less-constrained), reconstruction-history-driven regime selection reduced prospective prediction error and robot-side physical interaction conflict. The results suggest that accumulated prediction mismatch can serve as an internal signal for adjusting reliance on learned dynamics.