As this PhD thesis aims to develop a collaborative virtual reality environment for “Jeu de Paume” (JDP), for training and learning purposes, we have to consider several scientific questions. The first question to address in the simulation of JDP for training purposes, is to ensure biofidelity. To do this, we first need to perform a physical characterisation of different elements of this sport, essentially the ball, the raquette and the court. It is also important to collect data on JDP players in order to analyse the movements during sport practice. The M2S team has already a great experience of collecting such data in many sports, including tennis. The rackets used in JDP are asymetric, the balls are 30% heavier than in tennis but with a similar size, the court is asymetric and closed. The court is designed to make the ball bounce on the vertical walls, and on the inclined roofs of the “galleries'' around the ground called “carreau”. Another particularity is the fact that rackets and balls are still made using traditional craft techniques. The following question, in order to enable distant playing, is to propose a dynamic model of playing able to circumvent the technical impact of delay due to network. Recent works have demonstrated promising performance of deep neural networks, especially Recurrent Neural Network RNN, to predict the end of a human motion knowing the beginning. This approach is promising but requires a huge amount of training data, for very different motions, which will be difficult to collect for this specific context. Thus, we propose to explore and adapt few shot learning techniques, while taking the context of the game into account.
Plus d’informations :
[Website INRIA]