Robots have always been able to perform complex ballets of motion with varying degrees of precision. Early examples, dating back to ancient Greek "miracle" machines, tended to be halting and jerky in their mechanically driven movements. Computerized motion control allowed them to reach super-human levels of precision, with smooth, flowing trajectories a ballerina would envy.
The current generation of robots are even able to adjust their routines for variations in positions of objects in the environment. For example, I saw several demonstrations at the 2009 International Robots, Vision & Motion Control Show and Conference, (R&V) ongoing at the Steven, Convention Center in Rosemont, IL, of robots able to reach into a bin containing many interchangeable parts piled randomly, select the unit easiest to extract from the bin, figure out the most convenient way to grab the part, and pick it up. Issues like units lying atop one another so that their images overlap in the robot's field of vision, which just a few years ago were application breakers, are no longer an issue.
These robots, however, are only able to perform because a human wrote a detailed instruction program for them, which specified each motion in minute detail. These robots do not think or plan at any sophisticated level. They act as a human engineer has programmed them ot act.
Figure 1: Lipson's self-learning robot has to experiment to develop its own self image.
That's not how Earth's most sophisticated machines -- living animals -- operate. I did not come with an instruction manual. Neither did you. Neither did your family dog, or the fleas crawling on his back. Amoebae, which rely on hundreds of flailing scilia to get from where they are to where they want to go to find their next meal. A newborn giraffe has to figure out things like how many joints it has and how to use its muscles to stand up, and it has only until the next lion attack to solve the problem.
In an effort to learn how newborn animals do it, Cornell University Associate Professor Hod Lipson is studying what he calls "artificial life." That is, self-organizing systems whose only goal is to get up and move about on their own.
I ran into Dr. Lipson at his booth tucked into a corner booth in the Emerging Robotics Pavilion at R&V, next to headliner Toyota's Partner Robot, which wowed audiences by playing trumpet (quite well, thank you very much). Lipson quietly pointed out that his robot, a four limbed star-shaped entity that looks like a cross between a starfish and a Lego set, is at the opposite extreme from Toyota's robot.
Partner is smooth, polished, and expressive, with fluid movements and a sophisticated reportoire of behaviors, which it received directly from human programmers who choreographed every move.
Lipson's unnamed entity, on the other hand, exhibits all the style and grace of a young ox. That's because it's on its own. All it started with was the knowledge that it had eight motors and two tilt sensors, and it wanted to get up and move. A video on display at the booth documents the stages it went through in its quest for self-image and locomotion. Starting with random motions it used to find what actuator motions caused what changes in tilt-sensor outputs, developed a self-image that accurately reflects its body's topology: four limbs consisting of two segments each with actuators controlling each joint, all connected to a central platform carrying the tilt sensors. Then it experimented with coordinated motions in an attempt to find gaits that would allow it to move about in various directions. Finally, it crawled off into the sunset (represented by the edge of the table).
Finally, Lipson simulated an injury by dismounting part of one limb. The movement it had learned no longer worked, so the robot had to, through trial and error, find another gait that would again allow it to walk off into the sunset, albeit with a decided limp.
Lipson's work obviously is helping life scientists understand what information an organism needs to have coded into its DNA to live and thrive. In addition, it may help future robotics engineers develop robots that can learn on their own, instead of needing detailed programming for every movement they make.
Leave a comment