We need practical and formal design methods for building integrated perceptual robots. A robot is, typically, a hybrid intelligent system, consisting of a controller coupled to its plant. The controller and the plant each consist of discrete-time, continuous-time or event-driven components operating over discrete or continuous domains. The controller has perceptual subsystems that can (partially) observe the state of the environment and the state of the plant. Vision as a passive sensing system is cheap, reliable and biologically validated, so we are pushing the use of vision for mobile robots as far as we can.
The structure of the robot follows standard models [1,22] that decompose the robot into sensing, reasoning, and action subsystems, each realized at a hierarchy of scales. The finest scale handles control loops with a 100 Hz rate and a 10ms time horizon, and operates synchronously. Each coarser scale reduces the rate by a factor of 10 and increases the time horizon by a factor of 10. At the highest level, the time horizon is on the order of 10s of seconds, and the system operates asynchronously. In practice the data flows up and down the time and space hierarchy are implemented as streams of messages passing asynchronously through the system.
The finest level is also responsible for reaction to stimuli such as looming objects, our ``knee-jerk'' reflex, which avoids approaching objects, in a dynamic environment, where transport of information to high levels would delay response unacceptably.