Autonomous navigation
The study of the cognitive processes involved in a navigation task and their implementation on mobile robots naturally pose the problem of how a robot can learn to move in its environment. Our robot must behave like a rat to explore its environment in order to discover interesting places (providing a reward signal). We model a certain number of structures involved in this task (the hippocampus, the entorhinal cortex, the prefrontal cortex and the basal ganglia) in connection with a simplified model of the visual system and the system for processing idiothetic information. This task takes into account the following constraints: main information coming from the vision, no ad hoc marking of the environment (bitter or traces on the ground), no predefined metric map given to the system, and generally no information a priori on the environment. The resulting models are implemented on different mobile robots, and allow on the one hand to replicate experiments carried out in rats and to compare the results obtained with those observed in animals and on the other hand their performances are evaluated on scenarios of more complex navigation: in indoor (multi-room) and outdoor (both on and off-road in collaboration with the VEDECOM institute) environments.
However, thinking about the cognition of an isolated agent while ignoring the internal state of the system and the influence of social interactions on it can be a big mistake. Indeed, in addition to the communication functions, emotions can play a key role on several levels in the functioning of a robotic control architecture. In order to increase the autonomy of the developed robotic systems, our research also focusese on the mechanisms allowing a robot to self-evaluate its behavior via a neural mechanism for evaluating the prediction of its sensations and / or sensorimotor elements used by each navigation strategy. This continuous evaluation of the internal state of the robot (novelty, progress or regression) then leads to the emergence of an emotional state (for example Frustration). Different experiments conducted on real robots have shown how the neural activity of the system characterizing the internal state of the robot (motivational system, exit from the self-evaluation system) can simultaneously neuro-modulate the perception and selection of navigation strategies, as well as facilitate the learning of robotic behaviors and human-robot interaction.
Current theses
Approches neurorobotique intégrées pour la localisation et la navigation d’un véhicule autonome
Sylvain Colomer ( Prof. Olivier Romain, )
Robust localization and navigation of an autonomous vehicle: A neurorobotics approach
( Prof. Olivier Romain, )
Succeeded theses
From metacontrol to emotions: An application to robotic navigation in big environments
( , )
Addressing robotic navigation in big indoors and outdoors is a way to challenge our systems. The idea is that scaling up is not only a matter of more computational power. Instead, an autonomous robot should be able to self-evaluate, self-regulate and adapt its own behavior in order to cope with a dynamic unstructured environment. For that purpose, modeling internal and external signals — such as motivations, prediction errors, emotions, social interactions, contexts, etc. — is essential. This also allows for using robots, as artificial forms of life, to investigate what and how different kinds of information affect the way living beings can perceive and act in their physical and social environment.
Toward a plausible model of action selection for a mobile robot
( )
The main topic of the thesis is to develop bio-inspired actions selection models based on the emulation of the cortico-basal loops involving the hippocampus,the prefrontal cortex and the basal ganglia. The goal is to develop more robust robotic control architecture within the frame of computational neuroscience. By analyzing and reproducing the results obtained by neurobiologists, we hope to obtain a better comprehension of the brain functions allowing us to deduce models more suitable for real-world dynamic environments. The models describe functions involved in low-level action selections, like choosing to turn left or right; as well as in selection between high level strategies used in various navigational paradigms, like the competition/cooperation between sensory-motro habitual and high level cognitive strategie.
Bio-inspired visual navigation in large indoor environments
(, , and Dr. Ramesh Caussy)
The aim of this thesis is to ensure the scalability and the robustness of existing navigation algorithms in large and unknown environments (no prior knowledge).
Toward a hybrid and emergent approach for a adaptative-collaborative multirobot system based on ants behaviors
(, )
From self-evaluation to emotions: neuromimetic and bayesian approaches for the learning of complex behavior involving multimodal informations
Dr. Adrien Jauffret (, Pr. Philippe Tarroux, )
