Controle collaboratif entre un pilote humain et un module semi-autonome de navigation

The intervention of a human agent when a semi-autonomous navigation module is driving a mobile platform raises the shared control problem. This problem is especially complex when the semi-autonomous module and the human agent do not have the same perception system and do not avoid imminent dangers in the same way. The aim of this thesis is to develop a collaborative control approach between a human agent and a semi-autonomous navigation module by considering that the agents have different perception systems and different ways to avoid dangers.;By considering that the semi-autonomous navigation module helps the human agent during its navigation maneuvers, we propose a collaborative control approach based on an estimation of the human agent ability. The system consists of two main modules: a semi-autonomous navigation module which allows the direct integration of the human agent control signals and a deliberative module based on the estimation of the behavioral entropy of the human agent.;We have introduced the concept of the mechanical reflex arc defined as the association between a danger context and the human agent control signals. A new algorithm for online clustering of the mechanical reflex arcs is proposed, implemented and validated. When a motion deadlock occurs, the involved danger context is used to find a similar context in the set of available mechanical reflex arcs. The experiments in the laboratory have shown that the directional potential field approach combined with the use of the set of mechanical reflex arcs reduces the downtime due to motion deadlocks and produces a smooth and safe motion in a constrained environment.;The semi-autonomous navigation module support is effective if the human agent ability to avoid perceived dangers is taken into account. We show that this ability evaluation is linked to the human agent workload. We propose the concept of behavioral entropy estimation as a measure of this ability. This measure has the advantage of producing an instantaneous estimate without additional measurement devices. Using this estimate, we propose a deliberative scheme that solves the collaborative control problem. The analysis of this deliberative approach shows that it is effective in terms of speed of computation, and it covers the contexts of danger involved in this thesis.;The test results of the whole system suggest that the intervention of the semi-autonomous module when the human agent is navigating does not cause significant interference to the platform dynamics. Usually, the human agent is unable to say precisely the period during which he perceives that the semi-autonomous module helps in the execution of its navigation tasks. The compilation of the number of collisions shows that the collaborative mode is safer than the manual mode. In addition, the proposed collaborative control system allows the emergence of new platform dynamic behaviors such as the wall following and the doorway traversal.;The thesis main contribution to research is the design of a mobile platform collaborative control approach based on the instantaneous estimation the human agent ability. This approach extends the human agent ability to avoid dangers in a constrained navigation environment. The other innovative contributions are: 1. the development of a new deliberative scheme that is effective and fast and that allows both agents to safely navigate in constrained environments. This method does not require a direct exchange of messages between the human agent and the semi-autonomous navigation module. In addition, the deliberative scheme provides a smooth motion that takes advantage of the two agents perception system strengths; 2. the theoretical and the practical characterization of the ability of a human agent during the navigation in constrained environments. This measure, based on the real-time estimation of the human agent control signals safetiness entropy is called behavioral entropy; 3. the development of an algorithm for the inline clustering, which does not require the prior knowledge of the number of classes and the number of instances. (Abstract shortened by UMI.)...