Work Package: WP3
Deliverable: Deliverable 7.1: “Theoretical study of learning and adaptation in the dynamic field architecture for HRI”
Deliverable due date: month 18
This document reports on the progress of work on Deliverable 7.1. In the NETT 289146 Grant Annex, we stated in the context of Deliverable 7.1 and 7.2:
“P7 will analyse, numerically test and implement on the robotics system different neuro-plausible correlation-based learning rules and adaptation mechanisms. A direct objective of this work is to design autonomous robotic systems with human-like social and cognitive skills, with natural applications in health-care and real-world service tasks.”
In order to achieve the long term goal, it was necessary to gain knowledge and experience in several aspects of Human-Robot Interaction, with an emphasis on learning and adaptation. The ESR fellow focused on the theoretical framework of Dynamic Neural Fields (DNFs) applied by the group at UM in the areas of computational neuroscience and cognitive robotics. She developed, analyzed and tested in simulations different extensions of an DNF Model of Learning Sequential Tasks. The results have been reported in the article:
W. Wojtak, F. Ferreira, W. Erlhagen, and E. Bicho, “Learning joint representations for order and timing of perceptual-motor sequences: a dynamic neural field approach”, in The International Joint Conference on Neural Networks. IJCNN, 2015, pp. 3082–3088.
In this publication, we address the fast and efficient learning of neural representations that support the performance of precisely time action sequences. We ask specifically the question how feedback about executed actions might be used by the learning system to fine tune a joint memory representation of the ordinal and the temporal structure which has been initially acquired by observation. We propose and test in simulations a DNF model that detects a mismatch between the expected and realized timing of events and adapts the activation strengths in order to compensate for the movement time needed to achieve the desired effect.
The integration of the theoretical model in the control architecture of the humanoid robot ARoS has been finished. Currently, we are testing in different real-world human-robot interaction tasks the sequencing and timing capacities of the autonomous robot. A journal paper with the title “Adaptive timing for fluent human-robot interactions” (to be submitted to Frontiers in NeuroRobotics) is in preparation.
The remaining part of P7, that is, the final results of the “Demonstration of the robot’s learning and adaptation abilities” will be reported in Deliverable 7.2. The work also covers milestone 9.
Contributors: Weronika Wojtak, Estela Bicho, Wolfram Erlhagen, Flora Ferreira (UMinho)