Silicon Sidekicks - Intelligent Robots

Yelling "Fetch, Rover!" the man tosses a frisbee far across the yard. His dog just stands there and does nothing.

Oh yeah, he thinks, I forgot. "Rover! Turn clockwise until you face north-by-northwest, move forward ten-point-five meters and stop, lower your head fourteen centimetres, then close your mouth on the frisbee," he instructs. "After that ... return!"

The dog bounds across the yard, only to return with nothing in his mouth. The frisbee had been 2 cm farther forward than the man said - but the dog couldn't make that small adjustment on its own. So the man gives his long-winded instructions again, substituting "ten-point-five-two" this time.

But it could be worse. What if you were an astronaut newly-landed on the surface of Mars and Rover was your robotic assistant? With only a few precious months to explore Mars, Rover would be a tremendous waste of your time. Having no common sense or flexibility, Rover could even be a danger to you in the unforgiving environment of a distant planet. It might blindly follow instructions that cause it to damage vital equipment, and it would ignore the plain-language pleas of an astronaut in distress.

The Sojourner rover that explored Mars' surface in 1997 operated much like our clueless canine. Teams of scientists here on Earth had to feed Sojourner precise, step-by-step instructions for each task it performed. If the rover hit a snag, it would just stop and wait. The scientists then had to tell it exactly how to overcome the problem. It took days just to get simple tasks done.

Sojourner was nevertheless successful thanks to the ingenuity and patience of its controllers. Yet much more was possible. If we're serious about exploring the solar system, say mission planners, we must build smarter and more capable robots.

"During the next decade," says NASA Ames roboticist Liam Pedersen, "there's not likely to be a human presence much beyond Earth orbit. So if we wish to explore places like Mars, we'll have to send robots. No robots, no exploration. Period."

"Transmitting detailed instructions to essentially dumb robots is grossly inefficient and expensive - especially when there's lots to do," he adds. For example: Robots scouting Mars, perhaps in advance of human explorers, will reconnoitre vast areas. They'll sample hundreds of rocks, drill holes in search of frozen water, and take thousands of pictures. "If each of these operations takes several days and a standing army of mission controllers ... well, you can see how the cost increases."

In 1997, the Sojourner rover "sniffs" a Martian rock named Yogi.

The first humans on Mars will be just as busy as the scouts that precede them. Astronauts will have to set up the first base camp on an alien world and learn to survive in a place that makes Antarctica seem mild. And while they're at it, they'll collect thousands of measurements for scientists back on Earth.

"An astronaut's time will be more precious than edible gold," says Pedersen."They're going to need smart robot helpers."

How smart? The kind of intelligence that we usually take for granted in animals would do fine, says Pedersen. Animals effortlessly distinguish the objects in their environment based on the input of their senses. They can recognize threats, and they intuitively understand how objects move and behave. They can identify goals - like a little scurrying morsel of food - and then plan and perform all the actions needed to get it. And they know their own limitations of energy, strength, temperature, and endurance, and they're careful not to exceed these.

Credit: John Frassanito & Associates

An artist's concept of human and robot explorers working together on Mars.

Pedersen says, "try teaching this simple lesson to a robot: 'You can't turn a glass of water upside-down because the water will fall out.' To us, that's extremely obvious. It's common sense. But if you want a machine to understand that, you've got to spell it out in painful detail."

The computer brains of conventional robots operate in basically the same way as home computers do: They execute a fixed program of "if-then" logic and computations. The speed and precision of this approach makes computers extremely good at narrow, specialized tasks. But it also makes them inflexible. Confront a conventional robot with a situation outside the scope of its programming, and it's clueless about how to respond.

The adaptability and novel problem-solving ability of humans (and many animals) has proven very hard to reproduce.

Nevertheless, a patchwork of approaches to more-flexible computing has emerged. Among these are technologies like probability theory, evolutionary computing, natural language recognition and neural networks. Each provides a way to add learning or flexibility to a robot.

For example, scientists at Carnegie Mellon University taught a robot to steer a car autonomously for 98% of a drive across the U.S. - a project cleverly called "No Hands Across America." They first trained the robot by letting it ride along and watch as a human drove the car. The robot learned to associate certain visual inputs with the correct steering responses.

A simple example of a neural network. Input signals enter from the left, pass through the two processing layers, then emerge on the right as output signals. This architecture can perform surprisingly sophisticated logic, especially when feedback loops are added.

The "brain" of this robot was a computer simulation of a neural network, which mimics in a rudimentary way the architecture of animal brains. Input signals are processed by webs of "nodes" (neurons) and "links" (axons). Neural networks learn from experience and can associate general inputs with specific outputs: four legs + a bark (the inputs) = a dog (the output), for instance.

Pedersen cautions that the inner workings of organic brains are too poorly understood to mimic precisely. "While neural networks are in some ways similar to organic brains," he says, "they remain vastly less complex or capable."

Probability theory, especially Bayesian statistics, provides another path to machine learning, says Pedersen. It allows computers to operate not only in terms of black and white - true or false - but also in shades of gray. Machines that "think" using such statistical models learn well from new and unexpected experiences. ("This is where I would consider the excitement to be in robotics," notes Pedersen. "Watch out for an explosion in robot capabilities.")

Image credit: Grey's Anatomy

The human brain - we all have one, yet its inner workings are mysterious. Learning more about organic brains might help researchers program smarter robots.

Yet another possibility is evolutionary computing, in which computers "evolve" their own software. "Mutants" of an original program are tried, and those that produce better results are preserved. Their code is then mixed and mutated again - like sexual reproduction - to produce the next "generation," and so on for hundreds or thousands of generations. This software "evolution" can produce very effective problem-solving programs that are too complex for the scientists themselves to understand.

These and other novel approaches to computing form the foundation for smarter, more autonomous robots. Scientists draw from this toolbox to build into robots those abilities that we take so much for granted in ourselves: understanding the meaning of spoken language, figuring out all the little actions needed to complete a task, navigating across terrain and avoiding dangers - the nitty-gritty of autonomous exploration.

Progress is indeed being made. One prototype robot called Hyperion has shown the ability to autonomously traverse the terrain of the Canadian Arctic. Developed by researchers at Carnegie Mellon's Robotics Institute, this robot carefully navigates to avoid being caught in shadows, so that its solar panels are always receiving sunlight. And it's smart enough to know when it's lost or in trouble.

Another experimental robot, called the Extra-Vehicular Activity Robotic Assistant (ERA), is a true astronaut partner - roving on wheels side-by-side with a space-suited human. Scientists at the Johnson Space Center are using it to test advanced technologies like natural-language interaction and recognition of astronauts' gestures. Much of what they learn will help design similar assistants, not only for planetary surfaces, but also for Earth orbit and deep space.

Notes Pedersen: "Here at Ames we're working on a rover called K9 that will be able to do many things on its own. It can look at rocks, make measurements, and decide what's 'interesting.' K9 is a technology testbed for the 2003 Mars Exploration Rovers and for the 2009 Mars Science Laboratory (a.k.a. the Mars Smart Lander and Mobile Laboratory).

Other experimental robots are pioneering a different frontier: life onboard a spaceship. The Personal Satellite Assistant (PSA), for example, is a small floating sphere that can propel itself using fans through a spaceship's corridors. Created by Yuri Gawdiak and colleagues at NASA Ames, the PSA looks remarkably like Luke Skywalker's robotic light-saber sparring partner from Stars Wars. That's no coincidence, says Gawdiak, who dreamed up the PSA after watching the movie.

(left) An artists' concept of Robonaut working outside a spaceship. (centre) Yuri Gawdiak of NASA Ames and his Personal Satellite Assistant. (right) The smart rover K9 during field tests at NASA Ames.

The PSA will be able to do many things: talk to astronauts who want information from the ship's main computer; monitor the air (like a canary in a coal mine) for concentrations of potentially harmful gases, e.g., too much CO₂; or simply venture into situations that might be too dangerous or uncertain for their human crewmates. Such high-tech helpers would be welcomed on the International Space Station.

Other robots are best-suited for duty outside the spaceship. Robonaut, for example, is under development at the Johnson Space Center. It has the basic shape of a human - or rather a half-human. Its body stops at the waist. Its arms and hands are designed to be very dexterous, and its head contains video cameras. Astronauts, safely inside their ship, could perform routine maintenance or important repairs to the outside of the ship using Robonaut as a remote-controlled proxy.

If robots are going to live onboard spaceships, notes Pedersen, then the spaceships must be designed with robots in mind. "The need for this kind of system-level design - designing the robot and the spaceship to each suit the other - is often overlooked by non-experts," he says. The ship must have facilities for recharging and storing the robot, and the robot must be able to access the ship's computers and handle any necessary equipment.

The International Space Station and its robotic arm, Canadarm2, are an example of a well-integrated system. The arm crawls on the outside of the station - flipping end over end like an inchworm from one specially-placed handhold to the next. A custom-made trolley can quickly transport the arm from place to place when speed is of the essence.

Working together. Astronaut Jerry Ross floats above Earth, attached to one end of Canadarm2.

Canadarm2 is impressive, but like Sojourner on Mars it is neither smart nor autonomous. The arm moves only when commanded by a human.

The main reason for the gap in "smarts" between the robots in scientists' laboratories (like K9) and those that have flown in space is a lack of proven reliability. Pedersen explains: "The problem is that these advanced technologies do not have any flight history. Will they work under the demanding conditions of spaceflight? Mission managers are rightly conservative; they prefer to stick with well-proven solutions."

With time and field testing, though, the best among these technologies will prove their mettle - or rather, their silicon. Good thing, too, because future astronauts are going to want their silicon sidekicks.