AUTONOMOUS ROBOT
LEARNING
Can robots solve problems better than their programmers? In theory,the answer is yes. Equipped with suitable autonomous learning mechanisms, robots can learn by trial and error, and improve beyond the abilities of their creators. How could it be done in practice? This is the central topic of the Dragon project. In the Piraia project, predecessor of Dragon, we studied a robot self-simulator, executing in a physical robotıs control computer while the robot itself was operating. In some respects akin to human awareness, we showed that this ³robot awareness² is key to achieving fast and efficient autonomous learning. We verified the theory for a relatively complex robot locomotion task.
Our simulator was able to simulate the robot and its interaction with a static environment, which is adequate in a world uninhabited by other agents. However, in the real world, be it an office or a mine field, the robot is not usually alone, but must make out with one or more co-inhabitants. This means that the robot must be able to simulate other agents, as well as their interactions and awareness. The simplistic simulator that only cares about the robot itself and a static environment is insufficient. In other words, the robot must implement a kind of higher-level awareness. We are currently investigating this idea and its implications in the Dragon project.
AUTONOMOUS ROBOT
LEARNING
We study methods for programming and control of complex but fault-tolerant distributed real-time systems. Autonomous robots, i.e. robots that can manage on their own for an extended time, are examples of such systems.
Why canıt we always have an operator who supervises the robot? In some situations it may be impossible to communicate effectively with a robot, e.g. under water, or because the operator has become ill, and the robot must help her from an inaccessible location to a hospital. A process may run so fast that a human operator is unable to catch up. It could also be so boring that the operator simply doesnıt feel like monitoring it, or so complicated he cannot handle it.
In order to make a robot learn on its own, a robot can try different courses of action, evaluate the result of each trial, and memorize how well they work. Subsequently, it can concentrate on those alternatives that performed best. The evaluation criterion must be pre-programmed, but the details of finding a good strategy can be left to the robot itself.
The main problem of this learning strategy is that it takes a long time. The robot must perform all the trials physically, and it can only try one at a time. In order to find an optimal solution, the robot must search through a large number of possibilities, especially when there are a many degrees of freedom.
TAKING ROBOTS BEYOND STUPID
MACHINES...
The word ³robot² is commonly used for any mobile machine that contains a computer. In some sense, this is no different than just an ordinary machine. Even if it is controlled by a computer, we feel a lack of some fundamental property that distinguishes it from being merely a robot zombie.
One such property could be possession of a kind of self-awareness, or an internal software simulation model of the robot itself. For a simple machine, self-awareness would not be a very useful property. However, for a complex real-time system such as an autonomous robot, there are several advantages.
Self-awareness in trail-and-error learning allows the robot to avoid performing all possible trials physically. It is enough to simulate them. A fast simulator can search through many more alternatives. The robot can also run parallel simulations by dividing them between different processors. The simulator is a simulator in a general sense. For some problems, it may be enough to implement the simulator as a small set of if-then rules. For other problems, the simulator may need to solve differential equations.
...AND TAKING ROBOTS BEYOND
STUPID ROBOTS
In a co-inhabited environment, a robot will encounter agents whose behaviour cannot be extrapolated by simple Newtonian mechanics. The simple simulator must be extended to include simulation of other agentsı awareness, i.e. simulate their simulators, for reasonable predictive quality. A crucial problem is how to acquire a model of the other agentıs awareness. This is essentially a learning problem.
There are two ways of obtaining a model of an other agentıs awareness: One is by observation, and another is by communication. In the former case, the robot draws conclusions about the other agent based on sensory data. Although this may be a workable approach for some properties, it is difficult to convey abstract information, such as the other agents awareness, which is essentially a simulation program. The potential for transmitting such information lies rather with the latter method, communication.
It seems that an advanced level of awareness in an animal is strongly correlated with its ability to communicate. We conjecture that an important condition for robots to be able to co-operate is their ability to communicate, i.e. their ability to speak a Robot Interlingua. Such a language must contain sufficient expressional power to form abstractions from basic elements. It must allow robots to convey and bootstrap simulation models, without any other means than the communication channel.
Empathy and several similar concepts, which we consider typically human, appear to originate as simulations of simulations. Higher levels of simulations are relevant, for instance in game playing or other strategic situations. A robot that is able to perform such high-level simulations of simulations may be able to display behaviours previously thought to be genuinely human.
THE AUTONOMOUS LEARNING SICS
LABORATORY
The Autonomous Learning SICS Lab currently employs two persons, Martin Nilsson, full time, and Bjorn Levin, part time. The lab is located in Kista, near Stockholm in Sweden. We strive to build a solid theory in connection with adequate experimental verification. Thus, we greatly welcome comments, suggestions, discussions, and visits by pure theorists as well as dedicated experimentalists.
