S7 Artificial and robotic systems

Topic

This artificial perception system implements physical principles analogous to those of biological systems, albeit using different technologies and operating ranges. A digital camera serves as the most direct artificial analogue to an eye; its fundamental limit includes shot noise—stemming from the statistical and discrete nature of photons striking each pixel—similar to the limit established by another model in this set regarding the biological retina. High-sensitivity scientific sensors can achieve very high quantum efficiency and detect signals corresponding to very few photoelectrons. When a pixel receives a larger number of photoelectrons, the signal-to-shot-noise ratio improves roughly in proportion to the square root of that number.

LiDAR, in contrast, acts as an artificial analogue to the echolocation modeled elsewhere in this set; it emits a laser pulse toward an object and measures the time it takes for the reflection to return. Its angular resolution is limited by factors such as diffraction and optical aperture, while its range resolution depends on the precision with which the time of flight can be determined. Greater precision in the pulse and timing measurements allows for the detection of increasingly minute differences in distance, making LiDAR a key technology for the three-dimensional perception capabilities of autonomous systems.

Radar employs electromagnetic waves, and its range depends on factors such as transmitted power, antenna gain, propagation losses, and the object's radar cross-section. Because the received signal level depends heavily on distance, significantly extending the range may require substantial increases in power or other system enhancements. Synthetic aperture radar, however, leverages the motion of its platform to coherently combine observations taken from different positions, thereby synthesizing an aperture much larger than the physical antenna itself. Increasing the effective synthetic aperture allows for improved spatial resolution, within the system's operational limits. The fusion of an inertial measurement unit and a satellite positioning receiver using an extended Kalman filter constitutes a common method of autonomous navigation. The inertial unit provides high-frequency information, yet its error drifts progressively over time—a behavior comparable to that of the vestibular sensors modeled elsewhere in this set. In contrast, the satellite receiver provides an absolute position reference, typically at a lower frequency and subject to its own noise. The filter combines both sources to continuously estimate the system's state. As the inertial unit accumulates greater drift between satellite updates, a new external observation results in a more significant correction to the estimate, contingent upon the uncertainties assigned to each sensor.

Finally, a convolutional neural network can process images from this camera using an artificial hierarchy comparable, in certain functional respects, to the cortical visual hierarchy modeled elsewhere in this set. Convolutions extract local patterns via spatially shared filters, while successive layers build increasingly abstract representations from preceding ones. Thus, simple features can combine to represent structures, parts, and ultimately complete objects. Greater network depth and capacity can enable the construction of more complex representations, provided the architecture, training, and data volume are appropriate for the task at hand.

ID:4762