Fly Vision Chips, Caltech

Introduction:
In September 1999 Professor Christof Koch and Reid Harrison of the California Institute of Technology (Caltech) reported that they had built a silicon chip which was modelled on the visual system of a house fly. The chip, which contained integrated photodetectors, was successfully used to steer the heading of a small autonomous mobile robot.

Fly motion detection:
The brains of small flying insects are a favourite amongst neurobiologists and have been studied for over 30 years. Half of the fly's brain is dedicated to vision. The fly uses vision to stabilise its flight path, evade obstacles and predators, and to chase mates. It achieves all of this despite having 100,000 times fewer photoreceptors than humans.

Flight stabilisation is achieved though the use of visual motion information to estimate self-rotation. Back in 1956 a model of motion detection was proposed called "delay-and-correlate". In this model the intensity of the signal from a photoreceptor is compared with the delayed intensity from neighbouring receptors. If a signal from a receptor is followed shortly later by a similar signal from a neighbouring receptor, then this indicates motion in one direction.

This model has proven to be highly successful and has been used to explain characteristics of motion-sensitive neurons and behaviours in flies.

The vision chip:
The vision chip built at Caltech was modelled on the first 3 layers of structure in the fly's eye. The top layer consists of on-chip photoreceptors. These feed their signals to a second layer of "delay-and-correlate" circuits called "Elementary Motion Detectors" (EMDs). Each EMD then sends signals to another circuit called the "Horizontal System" (HS). The HS collects the outputs from all EMDs and sums them to produce output signals.

Each column of circuitry is 60x200µm in size. A single chip has 144 of these circuits spaced in a 6x24 grid. The chip works in real time and is implemented using analogue CMOS VLSI circuit technology. This means that all circuits operate with continuous voltage values and in continuous time with no clock. Because motion detection is performed locally each circuit runs at very low power (i.e. 5 microwatts).

The robot:
The vision chip was tested by being fitted to a small Lego robot (size: 13cm x 19cm x 22cm). The robot's task was to move forward in a straight line. This would have been trivial but for the fact that its two wheels were geared to rotate at different speeds. With no visual detection the robot always moved round in tight circles. When the vision chip was activated, however, the robot was able to correctly compensate for its mismatch in gearing and it successfully moved forward in a straight line.

Future work:
The research team at Caltech are now collaborating with a number of other universities to try and create learning algorithms for movement in the primate's visual system. They are also refining their fly vision detector and trying to give the robot a smaller body with a view to possibly building a robotic flying insect.

Links:
Caltech: http://www.klab.caltech.edu/analogvlsi.html