From: thinman@netcom.com (Lance Norskog)
Subject: TECH: mechanical animation in VR  %-)
Date: Thu, 20 Feb 92 14:50:23 PST



Rodney Brooks of MIT gave two lectures at Stanford this week.
His AI robot group uses a very different philosophy about AI
than standard approaches.  His stuff is intended to work in
real-time with very low processing requirements.  I think his
techniques can be use to build interesting reactive & moving
robots in virtual worlds.

He showed how his robot control systems work.  They're just dataflow
networks of nodes, where each node is a finite state machine with a
little control logic around it.  His example 6-legged walking insect
had 9 boxes (most standing for 6 copies of an FSM) and walked OK.  He
added 6 more boxes in 3 stages, creating better locomotion over
obstacles with each addition.

The 6-legged walker used feedback loops.  The robot that hunts for
an empty soda can and carts it off had many more boxes, but
no feedback loops at all.  Low-level things like "move the wheels"
would go out, and higher-level things watch these and countermand
them.  The motor control buses are lines of countermand boxes with
higher&higher-level behaviors later in the chain overruling lower-level
commanders.  A higher-level commander can decide when to overrule.
(This no-feedback method yields a completely planar circuit
which makes VLSI compilation much simpler.)

One of his students said there's a book by John Connell of
MIT explaining the schema and design methods.  Brooks' group has a
compiler that creates the FSM networks from higher-level descriptions;
I don't know if he'll be willing to liberate it.

Lance Norskog
thinman@netcom.com