Research Goals: The Artificial Mouse

The artificial mouse is a joint EU project between biolgical and robotics groups. The main goal of the project is to understand the rat/mouse whisker system and its relation to vision from a biological point of view as well as synthetically by building a robot. During my PhD I want to use the artifical whisker system we have developed for categorization tasks. Major challenges are to identify the relevant features of the whisker data as well as combining it with visual input from the camera. The sensory input from the whiskers will be used to train artificial neural networks. So far we have used unsupervised learning of receptive fields to develop an optimal representation of the statistics of our input data (see Predicting properties of the rat somatosensory system by sparse coding ).

The Robot

I have equipped a standard Khepera II robot with two artificial whisker arrays. The whiskers consist of natural rat whiskers glued on capacitor microphones. Each whisker is thus a single sensor. Data acquisition is done on a laptop with a PCMCIA data acquisition card.

AMouse Figure: The AMouse with its artificial whiskers.

Discriminating Textures by Active Whisking

Discriminable textures

In a first series of experiments, we evaluated whether textures of different roughness can be discriminated by the frequencies elicited in the whiskers. On the upper figure to the left, 4 examples of frequency signatures can be seen , each belonging to a different texture. By simple computing the difference between such reference signatures and a test signature, textures could be classified.

It is important to know that the data was recorded in an active manner. The motor signal was recorded simultaneously. Only if the motor signal was used for cutting meaningful pieces of the stream of data, textures could be discriminated. For more details please refer to our publication.

Studying the Influence of Morpholgoy for Simple Whisker-Guided Behavior

So far, the artificial mouse has been used for experiments on the morphology of the whisker array. We looked for an optimal arrangement of the whiskers by comparing different morphologies as to how well the robot was able to avoid obstacles. We found that a different morphology than the one found in nature is optimal for this task.

Robot sensory vx. physical space The purpose of this experiment was to study the influence of sensor morphology on a simple behavior. As can be seen in the figure, whiskers extend beyond the solid physical dimensions of an agent. This is important for example in cat behavior. Among other functions, Cats use their whiskers to judge the width of openings. From the input of their whiskers they can decide whether they will fit through the opening. Such a behavior can be tuned on two levels: 1) by adjusting the morpholgy of the sensor such that it makes the decision simple and fast and 2) by learning of the relation between sensory input and opening width. In this experiment, we deliberately excluded learning to evaluate the influence of the sensor morphology on the behavioral performance.

Robot morphologies Different arrangements of whiskers with respect to their orientation towards the robot body were tested on an obstacle avoidance task. The robot was equipped with a stereotyped avoidance reflex elicited by the combined input to a whisker array.

Exploration densities All robot morphologies were evaluated on their coverage of the test arena on 20 runs. The same starting point is visible in the bright white bin in the lower left corner. To assess how well/evenly the arena was covered by the different robot morphologies, the density of explorateon was measured as how often the robot was found in a certain area of teh arena. This is depicted in greyscale code on the figures on the left. Bright areas were visited most often. From these densities of exploration it can be seen that the morphology of thw whiskers determined how well the robot was able to naviagate. Only morphology B managed to drive through narrow passages, indicating that its sensory and its physical space matched best.

For further details see my publications.

See AMouse Project Homepage,

2004-10-12

	The purpose of this experiment was to study the influence of sensor morphology on a simple behavior. As can be seen in the figure, whiskers extend beyond the solid physical dimensions of an agent. This is important for example in cat behavior. Among other functions, Cats use their whiskers to judge the width of openings. From the input of their whiskers they can decide whether they will fit through the opening. Such a behavior can be tuned on two levels: 1) by adjusting the morpholgy of the sensor such that it makes the decision simple and fast and 2) by learning of the relation between sensory input and opening width. In this experiment, we deliberately excluded learning to evaluate the influence of the sensor morphology on the behavioral performance.
	Different arrangements of whiskers with respect to their orientation towards the robot body were tested on an obstacle avoidance task. The robot was equipped with a stereotyped avoidance reflex elicited by the combined input to a whisker array.
	All robot morphologies were evaluated on their coverage of the test arena on 20 runs. The same starting point is visible in the bright white bin in the lower left corner. To assess how well/evenly the arena was covered by the different robot morphologies, the density of explorateon was measured as how often the robot was found in a certain area of teh arena. This is depicted in greyscale code on the figures on the left. Bright areas were visited most often. From these densities of exploration it can be seen that the morphology of thw whiskers determined how well the robot was able to naviagate. Only morphology B managed to drive through narrow passages, indicating that its sensory and its physical space matched best.