The realization of our cognitive architecture onto a hardware substrate is one of our main concern.
Here we describe a bio-inspired architectural approach to design highly adaptive systems in the context of mobile robotics. The concerned robots evolve in an indoor unknown environment and then exhibit several behaviours such as landscape learning, obstacle avoidance, path planning, sensorimotor control.
We aim at designing the intelligent embedded controller of those robots. The controller will be able to selforganize its elements in order to adapt its architecture to the robot behaviour.
An artificial retina is a device that associates an imager with processing elements onto a monolithic circuit. Yet, in the usual design of artificial retina, analog hardware will not suffice to process high-level cognitive computations. The most often, an on-chip array of bare processors is inserted in the retina matrix.
The SATURN architecture has been thought to benefit in the next years from use of three dimensional chip integration. Thus, we are working on a multi-layer architecture composed of:
[L1] an acquisition plan,
[L2] a memory plan, which memorize the primal visual information,
[L3] a communication plan, where global communication can take place between non-neighbor cells,
[L4] a computing plan, composed of computing cells which can communicate with their direct neighbors,
[L5] a learning plan, where the system tries to learn and adapt the configuration of the underlying layer according to the complexity of the visual scene.
This architecture is currently simulated or emulated on FPGA circuits.
We are now focusing on layer [L4]. The controller of the robot is implemented as a single system-on-chip embedded into the robot body. During its life time, the robot exhibits several behaviours that are considered as concurrent processing tasks for the controller.
The main goal is then to adapt the number of processing elements of a regular architecture to the current urgency of a task, where urgency is a function of the input-data activity (saliency). The adaptation mechanism is entirely distributed, thus the tasks placement onto the architecture resources is computed on-line by the system itself and does not require any external decision making (local or deported operating system).
The idea is to implement cells as reusable elements which behave as competing neurons in their basic configuration, but can be used by a task as part of its processing. So, each task is implemented as a set of cells. The set of unused cells can be colonized as additional resources to implement bigger tasks. The growing cell sets are thus restricted by the competition between their border cells. Since competition is computed from the activity of neighbour neurons (and from input saliency), the geometry of a task in the architecture is not limited to a 2D-rectangle as for standard reconfiguration, but by the number of competing tasks.
As depicted in the figure, the architecture of a cell is thus composed of: direct wiring with its neighbours, a local programmable data-path, a local memory storing program and data, and a configuration port controlled by the local neuron.
Four types of information are exchanged in the architecture: the input sensors data on which are computed the different tasks, the results of each processing broadcasted to the neighbours, the program of a replicating cell, and the activities of neurons.