Thursday, October 5, Volen 101, 2:10-3:10 pm. (Refreshments at 2:00pm)
Our hypothesis is that robots fail because of an underestimation of the complexity of the software design problem. Traditionally, engineers will build a complex robot, complete with powerful motors and sensors, show that a human with joysticks, or a programmed "script", can control it, and blame the software engineers for the failure.
In nature, the bodies and brains of creatures arise together, result of a long series of small mutual adaptations. There is never a situation in which the hardware has no software, or where a growth or mutation - beyond the adaptive ability of a brain - survives. This chicken-egg problem of body-brain development may be best understood as a form of co-evolution.
In our lab we have studied coevolutionary learning in games, optimization, language, and problem solving, and have recently applied it to robotics. Rather than seek an intelligent general purpose humanoid robot, we are focused on systems that can "design" special-purpose mechanisms and controllers to achieve specific tasks in specific environments.
Using evolutionary computation and physical simulation, our software generates blueprints and controllers for robots which can be fabricated from reusable sensors, effectors, and chips, held together with modular parts (like Lego) or rapid prototyping (robotically manufactured). This co-evolution of software and hardware together seems to capture some of the principles of living systems, namely, that design emerges without a designer, from the interaction of evolution and physics, and that complex lifeforms exploit properties of their own medium of construction. Of course, so far, our robots are pretty simple.
For more information see www.demo.cs.brandeis.edu
Host: Liuba Shrira