DETAIL-Interact: Dynamic Engagement and Tailored Interactive Learning for Advanced Robotic Systems

Descripción del proyecto

Avance en el aprendizaje y la interacción de los robots

La robótica ha avanzado en los últimos años impulsada por las innovaciones en electrónica y tecnologías digitales. Como resultado, los sistemas robóticos modernos han demostrado una serie de habilidades interactivas y de manipulación novedosas. Sin embargo, debido a la novedad de estas capacidades, se ha investigado poco para explicar los procesos subyacentes. El equipo del proyecto DETAIL-Interact, que cuenta con el apoyo de las Acciones Marie Skłodowska-Curie, pretende descubrir cómo los robots adquieren habilidades interactivas y de manipulación de la fuerza similares a las humanas. El proyecto se centrará en tareas de manipulación interactiva ricas en contacto, en las que se explorará la generalizabilidad robusta, la adaptación en línea durante la interacción y la representación de habilidades interactivas. También se crearán modelos generativos profundos para el aprendizaje por imitación fuera de línea, lo que permitirá a los robots aprender de las demostraciones humanas.

Objetivo

The DETAIL-Interact project, short for Dynamic Engagement and Tailored Interactive Learning for Advanced Robotic Systems, aims to demonstrate how robots master human-like interactive and forceful manipulation skills. Current research on robot manipulation primarily focuses on motion planning for contactless tasks. However, the more challenging domain of contact-rich interactive manipulation tasks has not been thoroughly investigated. Three aspects are crucial: interactive skills representation, robust generalizability, and online adaptation during the interaction. DETAIL-Interact seeks to surpass the state-of-the-art in interactive and forceful manipulation by enabling robots to learn from human demonstrations and exploration strategies. This goal will be achieved using Deep Generative Models (DGMs) for offline imitation learning, which will learn interactive motion priors and the representation of interactive skills. These models will handle high-dimensional demonstration data, including visual inputs, multi-channel force feedback, and robot states among others, essential for mastering forceful interactions. Subsequently, Monte Carlo Tree Search (MCTS) will be innovatively integrated into the receding horizon control framework, utilized for the iterative refinement of online movement through learning and adaptation. This method will optimize movement trajectories and applied forces based on feedback from each interaction and the learned interactive skills. Through this iterative process, the robot will be able to adapt to various interaction environments with differing stiffness levels. The iterative learning process will enhance the robots' performance in precision-oriented tasks, such as selective plant pruning, fruit peeling and cutting, and precise chemical powder grinding. Overall, the DETAIL-Interact project aims to enhance robotic capabilities in interactive and forceful manipulation tasks and is expected to push this topic forward to a higher level.