Apprentissage automatique de comportements socio-communicatifs d'un robot humanoïde par téléopération immersive

par Duc-canh Nguyen

Projet de thèse en Signal image parole telecoms

Sous la direction de Gérard Bailly.

Thèses en préparation à Grenoble Alpes , dans le cadre de Electronique, Electrotechnique, Automatique, Traitement du Signal (EEATS) , en partenariat avec Grenoble Images Parole Signal Automatique (laboratoire) depuis le 01-10-2015 .

  • Titre traduit

    Learning socio-communicative behaviors of a humanoid robot by immersive teleoperation


  • Résumé

    Socially assistive robot with interactive behavioral capability have been improving quality of life for a wide range of users by taking care of elderlies, training individuals with cognitive disabilities or physical rehabilitation, etc. While the interactive behavioral policies of most systems are scripted, we discuss here key features of a new methodology that enables professional caregivers to teach a socially assistive robot (SAR) how to perform the assistive tasks while giving proper instructions, demonstrations and feedbacks. The three main challenges of learning interaction by demonstration: we should (1) collect representative interactive behaviors from human coaches; (2) build comprehensive models of these overt behaviors and a priori knowledge (task & user model, etc.); and then (3) provide the target robot with appropriate gesture controllers to execute the desired behaviors. The collection data will be performed with an immersive teleoperation technique by professional caregiver - a pilot. It is so-call beaming technique where the pilot can be "embodied" in to the humanoid robot to interact with a human subject. At the same time, all of interactive behavioral signals such as speech, gaze, head and arm moverment will be gather. We have alreadly finished the beaming platform for collecting data. In the next, the experiment for gathering data will be performed. After that, the collected data will be converted to discrete events which is used to build multimodal interactive behavioral model with some methodologies such as Hidden Markov Model, Dynamic Baysian Network (the two methods were used successfully with Human-Human interactive data). Furthermore, some other methodologies can be applied such as Partially observable Markov decision process or Deep Neural Network in order to improve accuracy rate of the models. Also, the gestures controllers such as gaze, head, arm, speech controllers have been designed and evaluated adequately by human subjects. Recently, LSTM methods have been used to model interactive behavioral model for a humanoid robot to perform a collaborative task named "Put That There". Online vs. offline learning model are compared with conditional statistical model (such as HMM, DBN) and have been shown the effective of the LSTM models. The method is expected to generate effectively action events for the humanoid robot in the RL/RI test.