If asked, researchers will claim that divers, deep submergence vehicles (both submersibles and one-atmosphere suits), and ROVs are all mature technologies. By this, they mean that limited opportunity remains for any dramatic operational improvements. In contrast, AUVs are not considered mature technologies, and hence engender great excitement about future intervention possibilities.

As a tetherless, unoccupied vehicle, the AUV offers freedom from limitations imposed by a high-drag, foul-prone, strain-vulnerable, and penetration-limiting tether. Being unoccupied, it also relieves all involved from liability associated with exposing a human to undersea hazards. The dream is that AUVs will carry out preprogrammed missions with limited direct human intervention.

How is this to be achieved? The plan is that the AUV, through a combination of supervisory human control, artificial intelligence, and unprecedented sensor integration, can become the future underwater intervention workhorse. Such a work-class AUV would share the same goals as other underwater intervention systems: do something useful at the work-site that makes progress toward fulfilling whatever intentions the project needs.

The Work-Flow Model
The task-oriented model described in Figure 1 provides a basis for examining various system aspects. The same model, borrowed from Feedback Control Theory, was applied in my earlier article to divers, ROVs, atmospheric diving suits, and subs. As before, the actuator may be the point where changes are made, but before we can act, we need to know the status of the work site. Therefore, it is the sensor systems that start the process.

Sensors convert one form of energy to another - perhaps detecting acoustical energy from a sonar echo or converting light into digital electronic signals. The signal then travels to an interface with a processor. In other forms of intervention that processor is human. In AUVs, that processor may be either a human or a computer.

The processor then compares intent (or goal) with present condition, makes a decision, and sends a command signal to an actuator. In the case of a computer processor, signals passed among the sensors, processor, and actuators may travel only through the vehicle systems (wiring, hydraulics, computer processor, etc.). Otherwise, they may be transformed and transmitted through the water to the AUV's human supervisor. Hence, like other intervention modes, the AUV has its own way of exercising the functions depicted in Figure 1.

The AUV sensor packages are the same sensor bundles available for other underwater intervention uses. However, because the AUV will interpret information and operate without direct human intervention, the sensor packages are geared less for human orientation and more for whatever machine processor operates the system.

In a truly autonomous system, the sensor data transmits only through internal circuitry to a central processing unit - presumably, a programable computer capable of managing different missions. The processor receives the data converted to a language it understands (like analog to digital conversion), compares it to a preprogrammed intent, and generates the command signal that travels through the internal circuitry to an actuator. Following the command, the actuator alters the world around the AUV. Sensors monitor the results and feed the new changes back into the system.

Similar to the secondary loop where an autonomous human diver works in the sea and provides progress reports or requests for assistance to the surface, there might be a secondary loop between the AUV and a human supervisor. In this variation, the human provides high-level instructions - that is, the human operator defines (or redefines) the mission goal.

The operator does not, however, exercise direct control over all AUV actions. The AUV itself initiates detailed actions to execute intent. Expanding the initial work-flow model to explicitly accommodate this loop results in Figure 2, where the new feedback loop to the human supervisor provides an opportunity for the supervisor to alter intent. Note that the transmission medium between vehicle and supervisor may be different (seawater) than the internal vehicle path (vehicle circuitry).

Sensor System
Have you ever noticed in your car or truck (the later models that have an onboard computer diagnostic) that the computer sometimes misdiagnoses problems because a sensor malfunctioned? In most feedback control systems, the weak link is the sensor. Due to malfunction, a sensor may provide invalid information. Our friend, the ocean, is a harsh environment for instruments. If we want an autonomous vehicle to remain in its work environment for extended periods of time, we have to improve sensor reliability.

Presumably, though, this is an evolutionary problem, not one requiring revolutionary solutions. Cheaper sensors allow redundant sensor bundles, so that malfunctioning sensors can be identified by a "voting system" (majority rules: functioning sensors tend to provide similar signals, malfunctioning sensors provide odd signals). Improved ocean sensors do, however, remain one of our critical needs for further AUV development. Transmission Medium
You may have realized by now that the truly revolutionary thing about an AUV is that it is not directly operated by human control. The question might then be posed: Why don't we just drop the tether, and communicate through water with our autonomous little robot?

Obviously, the first objection to dropping the tether is that loss of nice power traveling down to the vehicle. We might reason, though, that portable underwater power packages should improve over time. We would further claim that such improvement needs only evolutionary, rather than revolutionary, progress. Further, land-based transportation's growing interest in portable electric movement will provide a substantial boost to this power evolution.

The rub, though, is a little thing called bandwidth. What we are inevitably looking at is: How fast can we transmit information through the transmission medium?

Data transference depends on how fast we can cycle a signal. In the digital age, that means how fast can we go from on to off, and then back to on again. Each such change is a little bit of information. If we transmit the data signal electrically through a nice hard piece of copper wire, we can switch very rapidly (high frequency). Radiating the same signal as a radio transmission through air, we cannot switch quite as rapidly as through a copper wire, but still really fast - up to a couple of hundred million cycles per second.

Unfortunately, electromagnetic waves, which include radio waves, do not propagate well through water. Instead, we use sound. Acoustical waves propagate well through water up to about one hundred thousand cycles per second, three orders of magnitude slower than electromagnetic propagation through air. This means that, all other things being equal, we can transmit data through water only about 0.05 percent (one-two thousandth) as fast as by radio waves. For high data requirements like photos and video, this presents severe limitations. To keep the operator informed of what the sensors see and be able to process actuator command signals in real time requires data transmission speeds unavailable in this bandwidth environment.

Of course, we are in the age of signal black magic. A little thing called data compression is helping us say more in fewer on-offs. Figure 3 is a simple illustration of data compression. Notice we have a representation of a mermaid in black and white. One method of photo representation, called a bitmap, identifies a coordinate system for each microscopic dot that makes up the picture. Each dot has its own address. A bitmap assigns a color to each tiny dot's address. In this case, there are only two choices: black or white. Printing each dot or projecting each dot on a screen, then, reconstructs the picture.

If we notice that the majority of the space in the picture is white, then we might realize that it could require less data transmission to just send the coordinate addresses of those dots that are black to the printer or display screen. Then, the receiving end would just assume all coordinates not included in the list transmitted are to be displayed (or printed) as white. Hence, the same information has made it to the receiver with less data transmitted.

Such a simple illustration only gives a flavor to data compression. Sophisticated data alterations called fast transforms (fast Fourier, fast cosine, or fast wavelet transforms) not only make your head explode when you try to understand them, but they also make it possible to compress data to an impressive degree. Next time you are near your computer, try finding a bitmap changed to an identical JPEG (joint photographic experts group) format. Or, if you have the software, transform one yourself. The data reduction achieved in a JPEG file versus a bitmap is on the order of one-tenth. There is some loss of detail, but most would not notice the loss. The compression ratio is significant, but it does not fully compensate for the restricted bandwidth.

The basic idea in data compression is that we create new "languages" that say what needs to be said in fewer "words." Then, we engage a computer to translate the raw data into this more succinct language and transmit that through the water. On the other end, another computer processor translates back into useable format.

Of course, another possibility is that the technical experts in this area could figure out new ways to increase the actual through-water bandwidth through some acoustical channeling magic. Perhaps, too, someone might come up with a whole new transmission form not necessarily based on acoustics.

In either case - data compression or bandwidth increase - we are talking revolution, not evolution.

Interpretation
While we wait for mathematicians and other magicians to invent languages with which we can say more in fewer signal words (or paths with greater bandwidth), we could come up with another idea. Instead of having our onboard computer spend its time transforming data into compressed streams, perhaps we could just have it take over interpretation and actuator control signal generation. Great! Now we just keep all the signals internal in the underwater vehicle, and no stingy acoustical bandwidth is going to slow us down.

Here's the rub: Humans come preprogrammed to deal with surprises. Machines do not.

Just because computers excel at only two of our biological brain activities, we are tempted to think that computers are superior. Computers are computation-based, and computer memory is address-based. An address-based memory gives the computer perfect recall, but it provides the computer with no reference for importance. Human brains are more complex, and at the very least correctly identify patterns much better than computers. Also, our memories are associative, rather than address-based, providing us with value reference for any particular memory.

What does this mean? Take a look at Figure 4, which is obviously a picture of two cubes. Identifying that correctly is very easy for you. For a computer, this simple task is tough. Pattern recognition and associative memory allows you to immediately identify the similarities and the associations between the two cubes. A computer does not make memory associations, nor does it readily see similarities. Alter the angle at which the cube is viewed, and, to a computer trying to interpret a video screen, a cube takes on completely different geometric characteristics. To a computer, the object on the left is dramatically different from the cube on the right.

Imagine trying to write the computational program to get a computer to correctly recognize a valve from any three dimensional point of view, while viewing the valve in conditions of low contrast and floating debris, and there being some unexpected marine growth or corrosion encrusting the valve. Surprise!

Just how powerful is pattern recognition? Consider that chess is a very computational game. Movements are precise, tactical outcomes can be easily calculated by assigning values to pieces. Rules exactly define allowed actions, and the possible consequences of a series of actions (several moves) rapidly increase to astounding numbers. Chess is a seemingly ideal structure to showcase a computer's ability to accurately predict and score millions of possible outcomes; hence, quickly conquering any human opponent. The reality is that it took decades of concentrated software effort and a computer specially built to play chess to finally result in Big Blue, which could consistently triumph over human chess masters.

Humans rely on pattern recognition from viewing strategic placement of positions and a much less accurate memory, but one based on value judgments. With that, people beat generations of computer chess efforts at a very computational game.

The same kind of processing works to the advantage of humans in initiating actuator signals. While we easily develop neuromuscular patterns to coordinate even the most complex actions - whether directly through our on limbs or through control interfaces with machines - computers have to compute every move. Feedback correction for us is easy, all part of that pattern recognition. Computational corrections, devoid of pattern recognition, result in a monumental effort.

A report from the National Research Council, Undersea Vehicles and National Needs (1996), states: "In the future, AUVs should be capable of pursuing tasks with abstract descriptions. . . with the ability to replan and reconfigure the mission based on a wide range of changing internal and external factors".

These replan and reconfigure decisions are the challenges faced by a computational machine that struggles to recognize patterns and associate value with memory.

Critical AUV Needs
AUVs have passed the threshold where their onboard intelligence can navigate well enough to perform survey and sampling missions. However, the intervention AUV remains something lurking in our best imaginations.

By combining global positioning systems (GPS), acoustical sensors, and inertial guidance, a survey AUV knows its immediate location. Sonar soundings and imaging then allow them to gather data. With large memory storage, they retain that data until they return to their human supervisors.

Supervisory human control, artificial intelligence (AI), and sensor integration are the paths to an intervention AUV. The AI/sensor integration and bandwidth issues call for revolutionary advances, not just evolution.