Mises en œuvre de Commandes Neuronales par Rétropropagation Indirecte : Applications à la Robotique Mobile
Patrick Hénaff · HAL (Le Centre pour la Communication Scientifique Directe) · 1994
The aim of this thesis is to use multilayer perceptrons and the gradient backpropagation algorithm to experimentally implement neural commands for wheeled and legged mobile robotics. After an introduction to multilayer perceptrons and backpropagation, followed by an overview of some applications from a literature review, we show that the main drawback of learning control by backpropagation is the use of a desired output (direct backpropagation). We propose a solution that backpropagates a criterion that mathematically expresses the robotic objective and its constraints (indirect backpropagation). This solution avoids the use of an inverse model of the robot or a reference behavioral model, which allows the online application of backpropagation, thus enabling adaptive neural control. The experimental validation of this technique is carried out by controlling the position and orientation of a non-holonomic mobile robot with two independent drive wheels. We show the feasibility of the method, in particular the ability of the network to adapt to the kinematic constraints of the robot. As a preliminary study for the neural control of legged robots, we apply the same method to the control of the dynamic equilibrium of a planar biped in simulation. We show that the off-line backpropagation of a criterion allows maintaining the dynamic equilibrium of the system without any a priori knowledge of the trajectory of the legs. When used online, learning significantly improves the control of the equilibrium