(Closed) MSc project: Deep Reinforcement Learning for Human-Robot Interaction

Duration: 6 months

Short description: This internship proposal is part of a broader research project where our goal is to provide robots with basic social skills, such that they are able to interact with human beings in a fluid and socially acceptable manner. With this long-term goal in mind, we designed and developed a reinforcement learning approach in which a robot can autonomously learn to adapt its gaze control strategy for human-robot interaction [ 1 ]. In this context, the robot learns to focus its attention on groups of people from its own audio-visual experiences and, to do so, our framework uses recurrent neural networks and Q-learning to find an optimal action-selection policy. As an important aspect of the learning carried out, we use a synthetic environment simulating moving people to pretrain our model and to avoid long periods of interaction with people in a real environment. The aforementioned research work could be extended by modeling other group behaviors such as speech turn-taking, or head-pose and gaze of the participants in a conversation or social meeting. In this scenario, we would need to automatically detect voice activity as well as to estimate head poses and gazes. We could then generate synthetic data following this group behavior model and consequently improve the decision taken by the robot. Another possible improvement of the existing approach could imply the consideration of each person’s identity by combining our model with a multi-person tracker. More precisely, instead of considering simple joint heatmaps as observations, we could have a set of joint heatmap (1 per person) with temporal consistency. This is a challenging issue since the dimension of the observation would be dependent on the number of people in the scene.

Keywords: Reinforcement Learning, Deep Learning, Human-Robot Interaction, Neural Networks, Multimodal Fusion.

Information for applicants: Please send your complete CV to Pablo Mesejo (pablo.mesejo-santiago@inria.fr)


[1] Stéphane Lathuilière, Benoit Massé, Pablo Mesejo and Radu Horaud Neural Network Reinforcement
Learning for Audio-Visual Gaze Control in Human-Robot Interaction , 2017, https://arxiv.org/abs/1711.06834v1 .