On-board architecture for the Autonomous Generation of Perceptual Learning

Principal investigator: Olivier Gapenne (BMBI)



The general objective of the AGPL project is to explore and characterize the conditions for autonomous generation of a learning strategy within a live agent (rat) and an artificial agent (robot). In other words, the challenge is summarized as follows: how does an agent, with action and capture means and a structure linking them, learn to perform an unknown task?


Methodology: The AGPL project consists of two complementary projects: the first one concerns the design of at least one control architecture of a terrestrial robot for autonomous generation of a learning rule (line tracking task); the second one is the study of learning mechanisms at work in a living rat holding an artificial perception system: rats have to learn to follow a line with a new system for perceiving a specific length wave, not initially perceived by the animal. The general vision is that behavioral biology inspires the functional architecture of the robot, and the robot architecture is ultimately an explicit model of the autonomous ability to produce a learning rule in a living rat.

  • Implementation and tolerance of an optical fiber in biological tissue:

Biological tolerance studies have been conducted to allow the intradermal implantation of an optical fiber of diameter 0.5mm. The reactions of the surrounding tissues were investigated by histological and immuno-cytological markers to verify the integrity of native sensory receptors and the interface with the fiber. The first objective was to prepare an implementation protocol of the mechatronic components on rats. We studied the possible dimensioning of devices with respect to the size of the animal. At the same time and to facilitate testing, a model was constructed using a LED, a phototransistor, a processing unit (Arduin) and signal viewing software (Processing IDE) to simulate current test conditions.

Then we had to select materials and cytotoxicity tests on biocompatibility of materials that we will use during implantation. MTS assay was chosen to determine the cytotoxicity of the materials involved in implantation in rats. Then for cytotoxic materials, we diluted extracts in different concentrations (five levels each) and made the MTS test, which helped us to quantify the trend of cytotoxicity of the material.

Internship: Rong Yi (2015/2016). “Implantation and biocompatibility of an embedded sensory system in rats”, under the supervision of Didier Gamet and Christophe Egles.


  • Study of the capture of a point light signal by a mobile optical fiber:

The first subtask aimed to design the generation of the path by laser scanning of a beam onto the surface on which the rat is moving. In the additional subtask, we suggested an optical localization of a mobile photo-detector installed onto the rat head, which could allow the localization of the rat on the exploratory surface.


  • Study of a perceptual learning ability in rats:

The aim of this subtask was to validate a situation for the behavioral study of the perceptive substitution in rats. On the basis of pilot studies, we characterized a situation in which a rat moves in complete darkness in search of virtual tracks indexed by a sound and leading to a water reward. We particularly described the evolution of learning during the different phases of the perceptive substitution and the strategies employed by animals. The data collected, including trajectories of rats, could be modeled for the purpose of understanding of this type of learning.

Internship: Sandra Gauthrin  (2014/2015). Study to validate a perceptive substitution situation in rats. Master Mention « Sciences Biomédicales, Spécialité Neurosciences et Sciences des Comportements, parcours Sciences des Comportements », Université de Caen, under the supervision of Vincent Roy and Olivier Gapenne.


  • Setting up a simulation environment to study the generation of a learning rule:

Testing and validation of learning rule for the robot require the completion of the operation of perception and action process, related by an action-decision feedback loop. The primary purpose of the deliverable is to study the generation of a learning rule, regardless of practical implementation issues. It is therefore proposed to design a simulation environment as accurate as possible in order to test the algorithms. These algorithms are the second objective of this deliverable. The community specializing in learning has many solutions that potentially meet the project needs. A solution will be identified and implemented in the simulator.

Internship: Adrien Fois (2015/2016), in progress. Master 2 on Embedded Systems and Microsystems, Université Paul Sabatier Toulouse, under the supervision of Jérôme De Miras.