Skip to Main content Skip to Navigation

Architecture robotique et cognitive pour l'apprentissage de tâches en interaction avec l'humain. Une application pour la collaboration homme/robot dans l'Industrie 4.0.

Abstract : Human-centric and flexible interaction in collaborative robotics is a key aspect of industry 4.0/5.0. Collaborative robots can now assist in many tasks, helping to reduce musculoskeletal disorders risks for human workers. However, the level of collaboration remains far from the natural one between two human coworkers. Indeed, reconfiguration of collaborative robots still lacks flexibility and is often out of reach of the everyday worker, who is neither a programmer nor a robotics expert. An ideal collaborative robot should become a Smart Robotic Assistant (SRA) that can adapt dynamically its behavior to the diversity of each situation, including tasks, environment changes, workers characteristics and their preferences. Such SRA requirements lead to a paradigm shift in the way collaborative robots are programmed.Throughout this thesis, to fulfill SRA specifications, we have explored the design of a prototype of cognitive architecture around the notion of Interactive Robot Learning (IRL). The robotic agent can be taught, by leveraging prior knowledge, how to represent and carry out unknown tasks with generalization abilities, according to workers preferences and characteristics. Teaching is done throughout interactions, in a mixed-initiative setting, incrementally, and in a fast and natural way by non programming experts.Taking inspiration from complementary AI and IRL paradigms found in the literature, we have highlighted the benefits of a hybrid architecture, interleaving symbolic and connectionist approaches. With SRA specifications in mind, we chose to develop a new cognitive system based on relational representations models and integration of specific learning modules based on deep learning. In particular, we have focused on exploiting modularity of behaviors representations for the agent deliberative and incremental learning process, which led to consider Behaviors Trees (BT) at the core of the behavior model. It helps to learn a hierarchical level of representations, from real world moto-perception to symbolic abstract representations.Experimental validations, with real collaborative robots, were made throughout the thesis to assess the behavior of the current architecture prototype with respect to our SRA specifications. As manipulation tasks are common in many industrial applications, we chose to focus these experimental validations on planar, task-oriented grasping scenarios. This has motivated the development and integration of specific based AI learning modules, leveraging humans demonstrations for learning grasping. From a few demonstrations, workers can teach quickly and naturally authorized and prohibited locations concerning the task and/or their own preferences.In addition and as future integration perspectives, we discuss how uncertainty and estimation techniques for deep learning could be leveraged in the core of the architecture, for failure predictions and active learning.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Monday, May 16, 2022 - 3:59:15 PM
Last modification on : Friday, August 5, 2022 - 2:54:01 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03669479, version 1


François Helenon. Architecture robotique et cognitive pour l'apprentissage de tâches en interaction avec l'humain. Une application pour la collaboration homme/robot dans l'Industrie 4.0.. Traitement du signal et de l'image [eess.SP]. HESAM Université, 2022. Français. ⟨NNT : 2022HESAE007⟩. ⟨tel-03669479⟩



Record views


Files downloads