HUGHIE, Louie and Heavenly sit on an enormous granite table in a basement, surrounded by an assortment of objects. Nearby, Jeff Russakow is looking at a picture of the table top on his computer. He highlights a small building block and moves it across the screen. Suddenly, Louie springs into action. Floating on a thin layer of air, the robot uses pneumatic thrusters to phut-phut its way towards the object. It grabs it and, phut-phut-phut, moves it to the new position.
The basement is in the Aerospace Robotics Laboratory at Stanford University in California, and Hughie, Louie and Heavenly are the forerunners of a new generation of free-flying space robots. The smooth, granite table allows them to float on a cushion of air, free from the normal constraint of friction. In this way, the researchers can simulate space flight in two dimensions.
In space, the robots will look rather like miniature satellites with arms. As well as thrusters, which rely on a finite supply of compressed gas, they will be equipped with momentum wheels, powered by electricity from solar arrays. Changing the rate at which the wheels spin will make the robots twist and turn in space.
Advertisement
The potential of the robots is clear. 鈥淚f you spend $600 to $700 million you can send three astronauts up on a Shuttle to go and retrieve some wayward satellite,鈥� says Russakow, a PhD student. 鈥淏ut it鈥檚 expensive and risky. For a few tens of millions you could send up a couple of disposable guys like these.鈥�
The actual construction of the robot is not the laboratory鈥檚 chief concern. At ARL they are interested in providing robots 鈥� whether they are destined to work in space, under the sea or on the factory floor 鈥� with enough intelligence for them to make decisions and plan. And Robert Cannon, the head of the laboratory, is in no doubt about how to concentrate the minds of his researchers: 鈥淣o one gets a PhD without doing an experiment that鈥檚 never been done before, and of course, it has to work.鈥�
The ultimate goal is 鈥渢ask-level control鈥�. Instead of controlling every move, a human operator will be able to say 鈥渕ove this to here鈥� and the robot will work out how to do it. 鈥淒ecision-making and planning is still beyond the state of the art,鈥� says Russakow. 鈥淚f you tell a robot to go to the store it can鈥檛. It can walk maybe, but it can鈥檛 navigate an unstructured environment where there are many objects that it hasn鈥檛 seen before.鈥�
Task-level control would be a valuable asset on the factory floor. Down the corridor, Steffan Sonck, another postgraduate, is using similar principles to Russakow to guide a pair of robotic arms in an experimental automated manufacturing station. He tells the robots how he wants objects fitted together and it is up to them how they go about it.
鈥淥ne of the big problems in the manufacturing arena is not so much that robots can鈥檛 do things. It鈥檚 that it is very expensive to change from one thing to doing the next,鈥� says Sonck. But all he has to do to alter the way in which his robots fit objects together is to draw a picture of what he wants on his graphical user interface.
Buzzwords
So what kind of programming is involved in developing these systems? 鈥淲e鈥檙e probably using every programming technique going,鈥� says H.D. Stevens, another ARL colleague. 鈥淭here are some places where we鈥檝e used neural networks or genetic algorithms 鈥� a lot of different artificial intelligence buzzwords.鈥� But ARL is not an AI laboratory. 鈥淲hat we do is look at what鈥檚 coming out of that field and ask, now how do I use it in real life,鈥� says Stevens.
Russakow鈥檚 robots illustrate how far ARL has progressed towards task-level control. When he drags a picture of an object across the screen, the instructions are radioed to the robot and the rest is up to its on-board computer.
The robot鈥檚 brain is programmed as a finite state machine. Every situation that it might find itself in is defined as a separate state, with stimuli such as an instruction or an impending collision causing the machine to transfer into a different state. 鈥淔or example, if you鈥檙e sitting around in a state of idle and someone tells you to move, you are pre-programmed with a response that puts you in the moving state,鈥� says Russakow. 鈥淚t鈥檚 not the most advanced technique but it is very powerful.鈥�
The robot鈥檚 control works on two levels. At the higher, planning, level, it might decide how to intercept and recover an object. The lower level controls the robot鈥檚 basic functions, for example, when to operate its thrusters to maintain a planned course.
The instructions stimulate the robot to transfer from the idle state to begin a series of checks. Is the message directed at the robot? Can it 鈥渟ee鈥� the object it wants to move? Is anything in the way? It then plans the best trajectory to intercept the object. Next comes the 鈥渕oving across the table鈥� state, followed by another state in which the robot coordinates its arms and grabs the object.
Each state requires a different set of control instructions. They can be thought of as software subroutines, with different stimuli causing the computer to shift between different routines. 鈥淎 human has decided all the crazy things that are likely to happen to the robot and programmed it with responses,鈥� says Russakow. For example, if a robot needs to compensate when an object it reaches for starts moving, the programmer must incorporate a 鈥渃ompensation state鈥� into the overall control.
But clever programming can鈥檛 solve one problem. How does a robot keep track of its position when it is flying around in space? At the moment, Hughie and Louie rely on overhead cameras to act as remote 鈥渆yes鈥� and watch the table. But in space there鈥檚 no ceiling from which to observe everything. Heavenly may have the answer.
Kurt Zimmerman has kitted Heavenly out with Global Positioning System receivers, which in space would monitor signals from the network of GPS satellites. The GPS satellites were launched by the US Department of Defense as a navigation system. By comparing the signals from three satellites, a GPS receiver can calculate its position on the ground. With four satellites, a receiver can pinpoint its position in three dimensions. The network has been used for everything from keeping trekkers on track in the jungle to monitoring the swelling of volcanic craters.
In order to get the precise accuracy necessary to prevent collisions when linking up with satellites in space, Heavenly relies on a technique called differential carrier phase GPS. Here the robot and the target satellite are equipped with GPS receivers. 鈥淭he satellite just collects its measurements and broadcasts them to anyone who鈥檚 around,鈥� says Zimmerman. The robot receives those and combines them with its own measurements to work out its relative position and attitude. You can figure out your relative position much more accurately than your absolute position in space.鈥�
In the basement laboratory, however, Heavenly has no access to the real GPS signals, so a set of antennas have been placed strategically around the room to broadcast signals that are identical to those from the real satellites. 鈥淓ffectively, we can receive from six different satellites right here on the table,鈥� says Zimmerman.
Flying start
The indoor GPS approach is attracting interest from industry. 鈥淵ou can use GPS indoors for sensing manufacturing robots or tracking large parts, such as half a plane, when they are being moved across the factory,鈥� says Zimmerman. 鈥淭he idea is to replace overhead vision systems. You can still use a local vision system to get more precise control at the end.鈥� Heavenly is equipped with two small cameras for close-up control during delicate operations.
But perhaps the most impressive demonstration of ARL鈥檚 expertise in GPS came in July this year, when the team won the Fifth Annual Aerial Robotics Competition in Atlanta. The competition requires a flying robot to locate six small metal discs scattered within a 2-metre ring. It has to capture them one at a time and ferry them to the drop-off point in a second 2-metre ring, all while flying without human intervention.
ARL鈥檚 winning entry was based on a small, heavily modified helicopter kit. Bruce Woodley, one of the design team, says: 鈥淭hese helicopters are very agile, which makes them a lot of fun to fly. But it also makes them very difficult, because everything happens so quickly. If you鈥檙e flying manually and you take your hands off the controls you have about two seconds before it crashes.鈥�
Typical civilian GPS receivers can give a position to an accuracy of around 100 metres, and even the military version only shrinks the margin of error to around 10 metres, which is not much good for flying small helicopters. Like Heavenly, the helicopter instead relies on differential GPS, this time with reference to a fixed ground station, rather than a target satellite. 鈥淏y comparing what we get on the ground with what we get on the helicopter, we can resolve the vehicle鈥檚 position to about one centimetre in three dimensions,鈥� says Woodley. And because the helicopter is equipped with four receivers, a quick comparison of their relative positions tells the helicopter what its attitude or orientation is.
The ARL helicopter was the only entry in the entire five-years of the competition to retrieve a disc successfully and move it to the drop-off point. Unfortunately, the disc retriever was a simple permanent magnet, so the helicopter could not let go of it once it got there. Maybe the team will have better luck next year.