麻豆传媒

IF THE wildest dreams of NASA鈥檚 planners come true, a probe will one day melt its way through the icy sheath covering Jupiter鈥檚 moon Europa. Once it reaches the sea that NASA believes exists beneath the ice, the craft will swim about searching for life before popping back to the surface to beam its findings home. And it will do all this by itself, without human intervention.

Even optimists at NASA admit that this type of mission is way beyond them at present. It鈥檚 impossible to anticipate all the troubles that might crop up while the probe is incommunicado beneath the ice. And the search for life could prove trickiest of all. 鈥淗ow would it even recognise life in an environment that鈥檚 so alien?鈥� asks Richard Doyle, head of the information and computing technologies project at NASA鈥檚 Jet Propulsion Laboratory (JPL) near Los Angeles.

But there鈥檚 a wider problem with the space agency鈥檚 entire exploration strategy. Instead of one launch a year, from about 2015 NASA has set itself the staggering challenge of launching one probe a month to explore the Solar System. If each has to be commanded from the ground, the signals between these craft and the ground teams will quickly clog the already over-stretched Deep Space Network, an array of communications dishes peppered across the globe.

Instead, Doyle and his colleagues at NASA have come up with a cunning plan: let distant craft make their own decisions. They aim to distil the knowledge and problem-solving skills of human experts into computer software, and let it take control. The software will act as a 鈥渞emote agent鈥� for the scientists hundreds of millions of kilometres away. 鈥淭he spacecraft will have to think for themselves,鈥� says Wesley Huntress, NASA鈥檚 associate administrator for space science.

Such spacecraft could also slash the number of scientists needed to run a mission. And that could mean huge savings, a big incentive for an agency always short of cash. NASA engineers will take the first step towards autonomous spacecraft this month, when they launch Deep Space 1 on a voyage to the distant asteroid 1992 KD. The most exciting part of its mission will not be the asteroid fly-past, but the host of new technologies it will test out along the way.

Deep Space 1 will be the first NASA spacecraft to rely on an ion engine. It will also have novel solar arrays, an integrated imaging system and a battery of other components that NASA hopes to use in the next generation of spacecraft. But perhaps most fundamentally of all, Deep Space 1 will be the first spacecraft to steer itself across the solar system using its own star map for guidance. What鈥檚 more, for six days Deep Space 1 will face the ultimate test of autonomy. It will have to think entirely for itself.

For this short period of the journey, a remote agent on the craft鈥檚 computer will assume complete control, overseeing the navigation equipment, thrusters, engine and other electrical and mechanical systems. And to give it something to worry about, researchers will tell the remote agent that the craft鈥檚 navigation equipment and thrusters have broken. If it solves these problems successfully, the software could be put in command on many future missions, including the probe to Europa.

At present, NASA controls its uncrewed craft by sending a stream of commands from Earth. Although they are fairly simple鈥攖ake a photograph, turn this switch on, for instance鈥攕ending these messages is a complicated, time-consuming task.

But with its revolutionary control system, Deep Space 1 should take much of the burden off the shoulders of the Deep Space Network and NASA ground control. The remote agent should be the perfect employee, says Barney Pell, one of its designers and an artificial intelligence specialist at NASA鈥檚 Ames Research Center near San Francisco. 鈥淵ou can give it a few words and you know that it understands exactly what you mean.鈥�

Bright spark

The agent software doesn鈥檛 rely on anything as fancy as a neural network. Instead, as a result of endless meetings between engineers and programmers, it has a complete knowledge of the spacecraft and a set of preprogrammed mission goals.

The agent鈥檚 job is so complex that Pell and his colleagues at Ames and JPL had to find a way to simplify the task. So they divided the agent into three distinct units: a planner, an executive and a fault diagnosis module (see Diagram).

It is up to the agent鈥檚 three parts to make sure the mission鈥檚 goals are achieved. 鈥淭he planner decides what should be done and provides the resources to do it,鈥� says Doug Bernard, the mission鈥檚 senior engineer. 鈥淚t looks at all the goals for the time period and asks: OK, how am I going to achieve these goals?鈥� adds computer scientist Ben Smith.

Once the planner has decided to snap a photograph, for instance, the executive carries out the command via a software interface. It may sound easy, says Smith, but simply firing a thruster can involve numerous lower-level commands.

The agent鈥檚 fault diagnosis module鈥攏icknamed Livingstone after the Victorian doctor and explorer鈥攕erves as chief engineer. Its job is to look for mechanical or electrical problems and, once it has pinpointed them, pass the information to the executive.

No one can hear you scream

Normally, when trouble crops up on a space probe, the flight computer turns off all experiments and noncritical electronics, and asks for advice from ground control. A craft controlled by an agent will work differently. Sensors will constantly monitor the health of the craft. If the fault-diagnosis module spots a problem鈥攆or instance if the solar arrays refuse to deploy properly鈥攖he planner should come up with a solution based on its expert knowledge of the craft鈥檚 electronic and mechanical systems.

Only if the agent is stumped will it alert the ground team to the problem. 鈥淲e always have a backup. If the planner can鈥檛 get out of a problem, it will get the spacecraft to contact the ground,鈥� Smith says. 鈥淵ou鈥檝e got to get it right.鈥�

Originally the agent was meant to run the entire Deep Space 1 mission, but various problems鈥攕uch as having to fall back on an older computer processor rather than trying out a new design鈥攆orced a rethink. Despite the launch being postponed for several months, the team decided there wasn鈥檛 enough time to test the agent fully before lift off.

Now the plan is to finish testing the software after Deep Space 1 is launched and beam it up to the probe as it flies toward 1992 KD. Even then, for most of the mission, the agent will be switched off and the probe will be controlled from the ground鈥攁lthough the autonomous navigation system will do most of the steering. 鈥淪pace people are conservative. Nobody wants to be the guy who tried something that failed,鈥� says Marc Rayman, the engineer in charge of the mission.

When the agent is finally switched on next May, Rayman will test it by setting it three problems. First, NASA will tell Deep Space 1 that a switch is stuck in the 鈥渙n鈥� position. To solve the problem, says Smith, the agent should try the simplest thing first鈥攆licking the switch off and on. 鈥淚t鈥檚 the most obvious thing to try,鈥� says Smith. The engineers have set things up so that this will solve the problem.

Next, the team will tell the agent that the camera can鈥檛 be turned off. This unit contains the detectors that will be used to study asteroid 1992 KD, as well as the camera used by the autonomous navigation system. The agent is programmed to shut the unit off between sessions to save power. 鈥淭he first thing it will do is reissue the command, but we鈥檝e set it up so that鈥檚 not going to work,鈥� says Rayman. With the power switch stuck on, the craft will use more energy than originally predicted.

This is where the planner comes in. It will have to ask: 鈥淗ow can I achieve the remaining goals given that I鈥檓 consuming more power?鈥� The answer, says Smith, might be to take fewer images with the camera, or to turn off other equipment to compensate. In this case, the problem won鈥檛 prevent it carrying out its goals.

Lastly, during a third test, the NASA team will simulate an attitude control problem. The agent will be told one of its thrusters isn鈥檛 working. The agent should then turn the faulty thruster off and switch to another. This will still be able to turn Deep Space 1, but it will be slower and less fuel efficient. 鈥淲e should see it switch to the new mode,鈥� says Smith. 鈥淚t will then need to account for this change in manoeuvrability and fuel use.鈥�

At each stage of the tests, the agent will send its plans to the ground. If things go wrong, Rayman and his team will be ready with special software that will give it the equivalent of a frontal lobotomy and put command firmly back in NASA鈥檚 hands.

Unlike the agent, which will be turned on for only six days, the autonomous navigation system will steer Deep Space 1 throughout the mission. Accurate navigation is obviously critical, but at present, to get a fix on a craft鈥檚 position and speed, ground controllers have to measure the changes in the frequency and direction of radio signals it sends back to Earth. Only when they know where the craft is can they tell the craft how to change direction. 鈥淭here鈥檚 an extraordinary amount of planning that goes into this,鈥� says Rayman.

There鈥檚 another difficulty too: just as a beam from a flashlight spreads out with distance, the farther a radio beam travels through space the broader it becomes. So as a spacecraft gets farther away, the increasing width of its radio signals makes it harder for Earth-based navigators to tell precisely where the craft is.

The autonomous navigation system should solve these problems. About three times a week during its nine-month journey, Deep Space 1 will use its camera to record the positions of stars and nearby asteroids (see Diagram). By referring to its very own star map, which has data on about quarter of a million stars, it will work out its position and, when the remote agent is turned off, the navigation system will communicate directly with the spacecraft鈥檚 software interface, controlling the engine.FIG-mg21575001.JPG

The autonomous navigation software package has never been tested in space, and the stakes are high. If the slightest error is introduced into its calculations, Deep Space 1 will rapidly become confused. And the challenge of navigating across the Solar System is compounded by two things: cosmic rays creating false images on the craft鈥檚 star photographs and small errors introduced by the craft鈥檚 xenon ion propulsion engine.

The solar-powered ion engine pushes the craft forward by ejecting xenon ions out the back at high speed (see鈥滱rtemis and the electronauts鈥�, 麻豆传媒, 21 October 1995, p 34). Though the thrust generated by the engine鈥檚 ghostly blue exhaust is tiny, over many weeks it can accelerate the craft to more than 13 000 kilometres per hour.

But small changes in the pressure and temperature of the ions as they leave the engine means that the engine鈥檚 thrust can vary from one moment to the next. This makes navigation tricky, says NASA engineer Steve Synnott. If the navigation system doesn鈥檛 take these changes into account, Deep Space 1 could end up millions of kilometres off course. So NASA have paid careful attention to monitoring and controlling engine performance during the mission. 鈥淭hat鈥檚 been a major complication,鈥� says Synnott.

But Rayman believes they鈥檝e overcome the problem. He and his team have developed a way of controlling the engine鈥檚 thrust that should correct the errors caused by changes in its output. And if that doesn鈥檛 work, small thrusters will put the craft back on course.

NASA must also take into account the high-energy cosmic rays that whiz about the Universe. Some of these particles might hit the sensitive CCD chips鈥攖he solid-state equivalent of photographic film鈥攖hat Deep Space 1鈥檚 navigation camera uses to record the position of stars. 鈥淐osmic rays can create what look like images of the stars and asteroids we鈥檙e looking for,鈥� says Synnott. To eliminate these false 鈥渟tars鈥�, the navigation system will compare images. 鈥淭he cosmic rays occur at random and will hit at other places in subsequent pictures,鈥� he says. 鈥淭he stars, however, will repeat in the same relative position over many frames.鈥�

Right on target

Since this is the first time a spacecraft has guided itself, NASA will keep a watchful eye on Deep Space 1. If it doesn鈥檛 get lost, in late July next year, the probe should fly just kilometres above the surface of the tiny asteroid 1992 KD.

It is here that the navigation system will really come into its own. After 188 million kilometres, and just three hours before closest approach, the system will take a couple of images to check its position and heading. If Deep Space 1 is on course, the navigation system will adjust the craft鈥檚 trajectory so that it skims just 5 kilometres鈥攁 hair鈥檚 breadth in cosmic terms鈥攁bove the asteroid鈥檚 surface. As it passes, the probe will take high-resolution photographs and study the shape and structure of the asteroid with its sensors.

If Deep Space 1 is a success, NASA will attempt to run an entire mission鈥擠eep Space 3鈥攚ith a remote agent and autonomous navigation in 2003, followed by the mission to Europa, scheduled for sometime after 2008. The whole future of space exploration next century might well depend on the brainpower of these remote agents.

Although Bernard and his colleagues are not yet working on software that can actually learn for itself, this could soon change. 鈥淯ltimately, we would like to have machines that can evolve their own intelligence,鈥� says Huntress. 鈥淓ven self-repairing machines.鈥�

But could the decision-making skills of the scientists at ground control ever be completely replaced by computers on spacecraft millions of miles away? 鈥淚 don鈥檛 think that will ever happen,鈥� he says. 鈥淢aybe I shouldn鈥檛 say never鈥攚ho knows what will happen in the next 100 years?鈥�