Embodied Robotics and Emergent Behaviors

Rodney Brooks heads the AI Lab at MIT, a pioneering center in embodied robot research. According to Brooks, by 2020 or so we will share the planet with robots that have emotions, desires, love and pride.One of their early successes was 'Genghis', an insect-like creature with six legs and compound eyes. Genghis’ eyes and legs are the inputs and outputs for simple behaviors such as 'chase', 'stand-up', 'walk over obstacles'. But when combined together in one body, cued by stimuli from the environment, the result is a robot that behaves like many insect predators we encounter in nature. Brooks describes it as having a wasp-like personality.Importantly, this was achieved with no central cognition. When independent observers witness Genghis, they can't help but describe Genghis' actions in terms of novel emergent behaviors for which Genghis has no programming or physical correlates.A reductive approach would deny this claim; if there are no correlates then the behaviors must be illusory. Even if we do not approach Genghis through the lens of reductionism there is a chance that our perception of its 'personality' is something we are projecting onto what we see based on our prior experience.

Let’s turn to another example. In this case the robot body has four wheels and a light sensor on the front that supplies power to the rear wheels in proportion to the light level it receives. (There is no digital computation on this particular robot.)

If you can run Java applications in your browser, here is a link to a software simulation of this robot: http://www.counterbalance.net/robot/robot.html (Note: on some computers this can take several seconds to initialize.)

With [Mode 1] selected, click [Start]. The robot - depicted as a Bumble Bee - will move forwards stimulated by the light.

It will go forward while there is light in the room falling on its sensor and stop when it has driven far from the light. This is machine-like behavior, and not at all intelligent.

However, if we add a second light-sensor, so there are now two where the Bee's antennae would be, and if we wire the left sensor to the right wheel, and the right sensor to the left wheel, what will happen? Just like a moth to flame, the robot will drive to the light from wherever it is in the room. If it incorrectly veers to the left, the right light sensor, seeing more light due to the angle, will automatically send more power to the left wheel and put it back on course. This self-correction occurs until it finds its goal. If the light is then moved, it will follow. The brighter the light, the more vigorously it will seek it.

To see this behavior in the simulation click [Mode 2] and then [Start].

By choosing [Mode 3] the robot finds itself in a field of many lights, and will automatically traverse the set, finding an efficient path amongst them all until the last one has been found.

How should we describe the behavior of such a robot? I think we must call it ‘light-seeking.’ Importantly, the robot responds in time, with what seems to be a beginning point, a period of trial and error, and a goal temporarily achieved, before potentially restarting its search if the light is moved. But there are entirely no programming or physical correlates to this temporally based behavior. A hardcore reductive approach would insist on describing this robot in terms of a dual set of light-sensors, motors and drive wheels. The light-seeking behavior would not be apparent if we reduce it to its parts and consider them in isolation. Since light-seeking functions are not apparent in the parts, any claim that it functions this way would be considered mistaken. I think this is obviously false. On the other hand, if the emergent behavior is real, would it be correct to say this robot has a goal - a telos - and an inbuilt disposition to achieve that goal, i.e. an 'intention'? It certainly behaves as if it does...

While Brooks is more than sympathetic to the reality of emergent behaviors, he believes that designing human-like robots will turn out to be relatively easy because "we are machines,""... nothing more than a highly ordered collection of biomolecules."I believe he comes to this conclusion by extrapolating his key insight that led to the success with Genghis and which continues with the Cog and Kismet projects. The insight was: leave out cognition. Prior to Brooks' work, the vast majority of AI researchers were trying to develop computer programs that mimicked human-like cognitive processes, and robots that used these kinds of programs to control parts of the robot by maintaining a high-fidelity software model of the robot’s state and the world around it. This turned out to be significantly harder than expected. Meanwhile, Brooks decided to see how far he could get by building robots equipped with just basic responses to their environment, and explicitly leaving out any large cognition feature. The answer was: surprisingly far.

Email link | Printer-friendly | Feedback | Contributed by: Adrian Wyard