This is a report about a modest attempt at endowing a mild mannered machine with a few of the attributes of higher animals.
An electric vehicle, called the cart, remote controlled by a computer, and equipped with a TV camera through which the computer can see, has been programmed to run undemanding but realistic obstacle courses.
Figure 1.1: The cart, like a card table, but taller
The methods used are minimal and crude, and the design criteria were simplicity and performance. The work is seen as an evolutionary step on the road to intellectual development in machines. Similar humble experiments in early vertebrates eventually resulted in human beings.
Figure 1.2: SRI's Shakey and JPL's Robotics Research Vehicle
The hardware is also minimal. The television camera is the cart's only sense organ. The picture perceived can be converted to an array of numbers in the computer of about 256 rows and 256 columns, with each number representing up to 64 shades of gray. The cart can drive forwards and back, steer its front wheels and move its camera from side to side. The computer controls these functions by turning motors on and off for specific lengths of time.
Better (at least more expensive) hardware has been and is being used in similar work elsewhere. SRI's Shakey moved around in a contrived world of giant blocks and clean walls. JPL is trying to develop a semi-autonomous rover for the exploration of Mars and other far away places (the project is currently mothballed awaiting resumption of funding). Both SRI's and JPL's robots use laser rangefinders to determine the distance of nearby objects in a fairly direct manner. My system, using less hardware and more computation, extracts the distance information from a series of still pictures of the world from different points of view, by noting the relative displacement of objects from one picture to the next.
A Mars rover is the most likely near term use for robot vehicle techniques. The half hour radio delay between Earth and Mars makes direct remote control an unsatisfactory way of guiding an exploring device. Automatic aids, however limited, would greatly extend its capabilities. I see my methods as complementary to approaches based on rangefinders. A robot explorer will have a camera in addition to whatever other sensors it carries. Visual obstacle avoidance can be used to enhance the reliability of other methods, and to provide a backup for them.
Robot submersibles are almost as exotic as Mars rovers, and may represent another not so distant application of related methods. Remote control of submersibles is difficult because water attenuates conventional forms of long distance communication. Semi-autonomous minisubs could be useful for some kinds of exploration and may finally make seabed mining practical.
In the longer run the fruits of this kind of work can be expected to find less exotic uses. Range finder approaches to locating obstacles are simpler because they directly provide the small amount of information needed for undemanding tasks. As the quantity of information to be extracted increases the amount of processing, regardless of the exact nature of the sensor, will also increase.
What a smart robot thinks about the world shouldn't be affected too much by exactly what it sees with. Low level processing differences will be mostly gone at intermediate and high levels. Present cameras offer a more detailed description of the world than contemporary rangefinders and camera based techniques probably have more potential for higher visual functions.
The mundane applications are more demanding than the rover task. A machine that navigates in the crowded everyday world, whether a robot servant or an automatic car, must efficiently recognize many of the things it encounters to be safe and effective. This will require methods and processing power beyond those now existing. The additional need for low cost guarantees they will be a while in coming. On the other hand work similar to mine will eventually make them feasible.<-- Previous  Next -->