aminat
ed areas. Those are definitely the trends that we at Carnegie Mellon's Robotics Institute are working to make happen. One other area is in everyday life, such as driving on the highway or in the home.
BYTE:
Where are we going to see robots showing up in cars and in homes? What will they do?
Kanade:
Well, in cars we are working hard to develop the capability of having a vehicle drive autonomously on the freeway. We have already demonstrated autonomous navigation based on computer vision in a car that was able to drive from Pittsburgh to San Diego, 98 percent hands-off, last year [1995]. Of course, I frankly don't know when people will completely take their hands off the wheel and let the robot or computer drive their cars throughout the journey. I would say it might be relatively soon.
But before that, you will see a more practical application, that is, cruise control more intelligent than the current one. The current cruise
control in cars is used just for speed control. The more intelligent one will change speeds relative to the curvature of the road. It will also determine whether or not you are in the correct lane or whether or not there is a potential obstacle in front of the car. The more intelligent cruise control tries to maintain the car in the lane depending on the environment such as distance relative to the car in front of it, the curvature, and the up slope or down slope, and control the speed based on that. And if the car is going off the way, drifting to the roadside, the system warns and even applies the brake. I would say that's probably short range.
BYTE:
When you say short range, do you mean five years?
Kanade:
I would say technically within two years, I am sure we are right there.
BYTE:
But then there's the process of selling it to the car manufacturers.
Kanade:
As well as acceptan
ce by the society. There is some psychological barrier there between current cruise control and the self-driving car that I am talking about. If you look at intelligent, anti-locking brakes, as an example, you can think of those brakes in one way as an intelligent robot. In spite of the fact that if you slam on the brakes when you want to stop quickly, the intelligent brakes are actually grabbing and releasing many times a second. This releasing of the brakes is, in the short term, opposite to your intention. But as a whole, the braking system tries to accomplish what you are trying to do, that is, stop the car. Now what I am talking about, the intelligent lane control, the ability to control speeds relative to the curvature of the road and so on, that certainly has a little more psychological barrier to be overcome by the drivers. But separate from that, I would say technically, I don't see any difficulty to reach the stage where we engineers say, "It's ready," within a couple of years, I think.
BYTE:
In terms of the Navlab 5 vehicle that drove itself across the U.S. last year, humans were still controlling the brake and accelerator on that one. But how fast could it go?
Kanade:
Well, we could go up to 80 miles per hour, though we don't do that on the public roads.
BYTE:
Today when people think of robots, they often think of those robots C-3PO or R2-D2 in the
Star Wars
movie series. But you appear to have a broader concept of what makes up a robot.
Kanade:
I don't think we need to limit our concept of a robot to that of a machine which walks or moves along separate from humans. You can think of it more as a symbiosis, or synergy, between machines and humans, which, collectively I call a robot. And here we have the human and machine cooperating. So in the car, you the human driver know where you have to go, and of course, over all, you are responsible to make sure the car goe
s where you want to go, its safety, and so on. But some part of the controls can be turned over to a machine. Sensors such as infrared or some microwave sensors that can actually see much farther and clearly than you can could tell you of potential hazards in front of the car. So that's a robot. It cooperates with humans, and results in a symbiosis for overall safety, comfort, and efficiency of our life. I think that's a very important viewpoint.
BYTE:
We already have machines similar to the ones you envision for automobiles in use in airplanes assisting pilots today.
Kanade:
Some airplanes can take off and land completely autonomously.
BYTE:
So what are the factors that are making these devices, or robots, more affordable for consumers?
Kanade:
The advancement of everything. Number one, sensors like cameras are now becoming so inexpensive that we can have systems that use many, ma
ny cameras, not just one. You can put a couple cameras looking in front, a couple looking in back, and more on the side. In that sense you have more eyes than yourself. Faster computers certainly help. In fact, the current Navlab Five is run off a laptop computer, instead of five SPARC 20 workstations it used to use. We used a supercomputer on the very first, the Navlab One. You see the advancement.
BYTE:
But in addition to more computing power, you are also improving your code, too.
Kanade:
Absolutely. I would say that the understanding of the algorithms is the most important part. It's misleading to say that the speed [of the computers] has advanced, therefore we can do this and that. The most critical part is understanding the problem itself, always.
BYTE:
So it's not just an issue of just throwing more MIPS at the problem.
Kanade:
Exactly. However, there is an interesting con
nection between the two, that is, the improvement of the code and the sped up computer. I would say that the fact that you can compute more [in a given time slice] certainly allows the vision researchers to think more systematically. Or to put it another way, we don't need to rely on a cheap trick just for the sake of making the algorithm run within a required time. We can think of a more systematic, more correct way to make things really work, rather than using a cheap trick to make things "appear" to work. There's a big difference.
BYTE:
What are the benefits of this more systematic approach?
Kanade:
The biggest benefit is we actually understand why a program fails when it fails. In the past you were often forced into a situation of "Well, we know this is what we really want to do, but the computer can't do all that in realtime." And then you look for ways to skip things, and areas to simplify. And when you do that, the designers of the algori
thms ourselves do not really understand when or where the effect of that simplification shows up.
Especially when it comes to building very reliable systems, it is very important to understand the effects of these simplifications. Now, due to better computers, we are relying less on having to use arbitrary simplifications just for the sake of making algorithms run faster. That in turn makes the system more robust and more understandable.
BYTE:
Your robots have explored volcanoes to gather data that previously researchers had to place themselves in peril to get. In fact, scientists have died in volcanoes during an attempt to capture samples. What are some of the other interesting ways that robots will be used to accomplish tasks that currently humans must perform at great risk?
Kanade:
Another one that we are working on is a helicopter project, a flying robot. This has an advantage in rescuing people in bad weather or fire fighting. It could
also be used in industrial maintenance, such as monitoring remote power lines.
BYTE:
You mean, checking to make sure power lines are installed properly?
Kanade:
That's right, but rescue is another very important scenario for us in bad weather.
BYTE:
How would the robot accomplish something that a human could not in that situation?
Kanade:
In bad weather, say for example, a ship sinks, and some survivors are expected. The first thing is to go and find them. Then when you find them, you can drop them supplies. I often hear in the news that the weather is bad and therefore the mission is on hold. Well, that's understandable. I would say, if the weather is good, then probably the ship would not have sunk. In these cases, the reason that the rescue mission is on hold in bad weather is that the people who would go out on a rescue would also be in danger.
Now if this is an unmanned m
ission, we can risk the failure. If the helicopter crashes, at least we are not losing another human life. On all of these dangerous missions, the key factor is, the more risk you take, the more effective is the mission. If you can actually do a task in such a way that even a crash is still acceptable, because no human life is lost, then I think the effectiveness grows rapidly.
BYTE:
How can it be more effective if what the robot copter is doing results in a higher risk of crashing?
Kanade:
In bad weather, finding a survivor in the sea, you have to fly low. And that's very dangerous. But the lower you fly, the better your chances to find them. But if you use a human pilot, you don't want to fly as low out of safety concerns, and this is less effective. So you have a multiplication factor. If a human flies twice as high [as a robot copter would], the mission is accomplished with relative safety, but then the effectiveness or possibility of findin
g survivors may be only half of what it would have been if you flew lower. Whereas if the robot searcher can actually risk a crash, then it can be sent to a height that's lower, and the probability to find them may increase. So, why not do it? There are lots of those cases.
BYTE:
And a robot helicopter could help fight fires too?
Kanade:
Yes, and again, the more effective the mission, the more dangerous. If you hit the hottest spot with a fire fighting agent, it's very effective. But it's also dangerous. So human pilots don't fly that low. The result is that many of the chemicals dropped are actually wasted. It's the same situation.
BYTE:
In terms of implementing these designs, bringing them out to the real world, what are the challenges for these robots to become more effective?
Kanade:
I would say the sensor data interpretation is the biggest challenge. Let us think of the robot'
s capabilities. A robot is useful because it is mobile. It is also useful because it can sense, including all kinds of senses, like vision, hearing, smelling, touching, and so on. It is effective because it can manipulate. It can drop something, manipulate something, move or dig, all kinds of things affecting the environment. Those three capabilities have to be realized. If you look at any robotics systems that have been successful, I think that so far, the success comes from the good combination of sensing, manipulation, and mobility. In some tasks, you may not have manipulation, you only have sensing and mobility.
The key among these three is the
sensing
at this point. Vision is definitely one of the most important aspects of that. You have to understand what is going on in order to react to the robot's surroundings. We have a long way to go before we can develop enough variety of sensing capabilities for all the tasks that you can think of. It's getting better, but that is always the challeng
e. Sensing includes both sensors themselves and algorithms.
The progress of interpretation methods is definitely slower than the progress of sensors themselves. Ordinary video vision is the example. Cameras have been pretty good for a number of years. In some cases, cameras are even better than the human eye. But the algorithms are far worse than human vision.
BYTE:
Are there any tools or technologies on the software side that will enable this to improve?
Kanade:
One of the good things is some of our algorithmic learning capabilities, including neural net, genetic algorithms, and multi-agent control methods. Those are definitely helping to achieve new capabilities. There are quite a few examples of those.
BYTE:
Such as?
Kanade:
One of the things that we did was detecting human faces in the field of view, where you point your camera at a scene and have a robot locate humans by
actually detecting the pattern of their faces. That program was not written in the sense of someone coming up with certain algorithms to do that. Instead, that capability was learned. It's important to point out, however, that these are not a blind application of neural nets. The problem of recognition, in this case, a human face, includes identifying the important features, what are not [important features], and how to present good examples of faces and examples that are not faces. In teaching face recognition, providing the face example is easy. The more difficult part is, what are the non-face examples? There are so many. And so we must decide what we should give the neural net as a non-face example. Should we give it a picture of a building? Should we give it a picture of a tree? A computer? It's an interesting question.
BYTE:
What about robots and entertainment, as in vehicles on the moon that people would control from Earth?
Kanade:
That's
a new breed of robotics, where we're getting into entertainment. In this case, the images of the lunar surface are sent back to the Earth and people can enjoy them as if they are on the moon.
BYTE:
So when the buggy goes over a dune or pothole, the floor underneath of the person back at Earth moves and shakes. Is that project still being financed by private money?
Kanade:
That's still on track, yes. Another project we have is virtualized reality. We have built a dome, called 3D-Dome. It's a five-meter dome where 51 cameras are looking inside. In other words, the dome space is covered by a sea of cameras. And whatever happens inside is modeled into a computer graphics 3D CAD model. So imagine an event inside the dome, and say a player swings a baseball bat. How his shirt, body, and hand moves, their 3D shapes, is modeled into the computer every 30 milliseconds. Since it starts from a real event, instead of artificially making it, we call it virt
ualized reality. Once we do that, we can place anything in that environment anyway you want. Thanks to the 3D models, you can actually immerse yourself into that environment. One of the best applications that I envision is what I named "watching NBA on the court." If you watch TV today, you are allowed to see only from the viewpoint that the director selected. And the camera is placed outside the court, because the physical camera would block the game.
BYTE:
So this would allow you to watch the game from the referee's point of view?
Kanade:
Actually from anywhere inside the court. Or if your living room is big enough, you can wear goggles and play with them. They don't know you are there, obviously [the actual basketball players], because you aren't really there. But you can see what it's like in blocking, say Michael Jordan. Or you can enjoy the viewpoint from the ball, if the ball could see.
BYTE:
There must
be other applications for this, such as physical therapy?
Kanade:
Yes, in education, physical therapy, or virtualized surgery. With virtualized surgery, medical trainees can see from any viewpoint, from the surgeon's viewpoint, from the nurse's viewpoint. Those are the kinds of new interesting things that we can expect to see. The important viewpoint I would like to convey here is that robotics is not limited to mechanical objects, such as arms and mobile robots. You have to look at it as an information oriented information intensive mechanism to help people. One example you can think of is, similar to the 3DDome, your house. If you have elderly people, your house may be filled with lots of sensors to monitor your health. This lets elderly people get more effective care while reducing the cost. In that sense, the house itself is a robot.
For more information on the Robotics Institute, see
http://www.ri.cmu.edu
.
Beyond R2D2: Robots Evolve
