Graph-based inner state prediction for a socially aware robot

• Autores: José Tejón Moreno, Juan Pedro Bandera, Antonio Bandera, Adrián Romero Garcés
• Localización: Proceedings of the XXIV Workshop of Physical Agents: September 5-6, 2024 / coord. por Miguel Cazorla Quevedo, Francisco Gómez Donoso, Félix Escalona Moncholi, 2024, ISBN 978-84-09-63822-2, págs. 103-115
• Idioma: español
• Enlaces
  • Texto completo
• Resumen

  • Robots working and interacting with people in daily life scenarios require to be aware of their context, both physical and social. Predictive capabilities allow the robot to adapt to its context, and are a key component of awareness. This paper presents a system that uses the current and past context knowledge to infer future context. This system is integrated into CORTEX, a cognitive architecture where immediate context is represented as an oriented graph, the Deep State Representation. The proposed approach is based on collecting and classifying all the information obtained from the executions of use cases performed in the past by the robot. Then it uses these data to train a Graph Neural Network to predict the next configuration of the Deep State Representation. The used neural network is a graph convolutional network. This kind of network uses the information of adjacent nodes in order to make predictions in the graph. The results presented in this paper shows that the proposed system is able to predict next states with a precision close to 100% for a known use case, using manually generated data for training, validation and test.