Cheaper than printing it out: buy the paperback book.

Out of Control

Behavior wants to be free, but to be of any use to humans, artificially generated behavior needs to be supervised or controlled. We want Robbie the Robot, or Bugs Bunny, to accomplish things on his own without our oversight. At the same time, not everything Robbie or Bugs could do would be productive. How can we give a robot, or a robot without a hard body, or any artificial life, the license to behave, while still directing them to be useful to us?

Some answers are unexpectedly being uncovered in a research project on interactive literature begun at Carnegie Mellon University. There researcher Joseph Bates fabricated a world called "Oz," somewhat similar to the tiny room of John and Mary that Steve Strassman created. In Oz there are characters, a physical environment, and a narrative -- the same trio of ingredients for classical drama. In traditional drama, the narrative dictates both characters and environment. In Oz, however, the control is inverted somewhat; characters and environment influence the narrative.

Oz is made for humans to enjoy. It is a fanciful virtual world populated with automatons as well as human-directed characters. The goal is to create an environment, a narrative structure, and automatons in such a way that a human can participate in the story without either crashing the story line, or feeling left out as a mere observer in the audience. David Zeltzer, who lent some ideas to the project, gives a wonderful example: "If we provided you with a digitized version of Moby Dick, there's no reason why you couldn't have your own cabin on the Pequod. You could talk to Starbuck as he went after the White Whale. There is enough room in the narrative for you to be involved, without changing the plot."

There are three frontiers of control research involved in Oz:

  • How do you organize a narrative to allow deviations yet keep it centered on its intended destination?

  • How do you construct an environment that can generate surprise events?

  • How do you create creatures that have autonomy, but not too much?

From Strassman's "desktop theater" we go to Joseph Bates's "computational drama." Bates envisions a drama of distributed control. A story becomes a type of coevolution, with perhaps only its outer boundaries predestined. You could be in an episode of Star Trek attempting to influence alternative storylines, or you could be on a journey with a synthetic Don Quixote confronting new fantasies. Bates, who is chiefly concerned about the experience of the human user of Oz, puts his quest this way. "The question I'm working on is: How do you impose a destiny upon a user without removing their freedom?"

In my search for the future of control from the perspective of the created rather than creator, I will rephrase his question as: How do you impose a destiny upon a character of artificial life without removing its freedom?

Brad de Graf believes this shift in control is shifting the goal of authors. "It's a different medium we are making. Instead of creating a story, I'm creating a world. Instead of creating a character's dialogue and action, I'm creating a personality."

When I had a chance to play with some artificial characters Bates developed, I got a sense of how much fun such personality petlike creatures could be. Bates calls his pets "woggles." Woggles come in three varieties: a blue blob, a red blob, a yellow blob. The blobs are stretchy spheres with two eyes. They hop around in a simple world of stepping-stones and some caves. Each color of woggle is coded with a different suite of behaviors. One is shy, one is aggressive, one is a follower. When a woggle frightens another woggle, the aggressive one stretches tall to scare away the threat. The shy one shrinks and flees.

Ordinarily the woggles hop around doing their woggly thing among themselves. But when a human enters their world by inserting a cursor into their space, they interact with the visitor. They may follow you around, or avoid you, or wait until you aren't around to harass another woggle. You are in the picture, but you are not controlling the show.

I got a better sense of the future of pet control from a prototype world that is somewhat an extension of Bates's woggle world. A virtual reality (VR) group at Fujitsu Laboratories in Japan took wogglelike characters and fleshed them out in virtual three dimensions. I watched a guy wearing a clunky VR helmet on his head and data gloves on his hands give a demonstration.

He was in a fantasy underwater world. A faint impression of a submerged castle shimmered in the distant background. A few old Greek columns and chest-high seaweed furnished the immediate play area. Three "jellyfish" hopped around, and one small sharkish fish circled the area. The jellyfish, in the shape of mushrooms and about the size of dogs, changed color depending on their mood or behavior state. Playing by themselves the three were blue. They would hop around on their fat monopod tirelessly. If the VR-guy beckoned them to come, by waving with his hand, they would excitedly bounce over, turn orange, and jump up and down like friendly dogs waiting to chase a stick. When he showed them attention their eyes would close in a happy expression. The guy could call in the less friendly fish by emitting a blue laser line from his forefinger and touch the fish from afar. This would change the fish's color and interest in humans, so it circled in much closer, and swam nearby -- but like a cat, not too close -- as long as it was occasionally touched by the blue line.

Even watching from the outside, it was evident that artificial characters with the mildest autonomous behavior and some three-dimensional form in a shared three-dimensional space had a distinct presence of their own. I could imagine having an adventure with them. I could imagine them as Jurassic dinosaurs and me really being scared. Even the Fujitsu guy ducked once when the virtual fish swam too close to his head. "Virtual reality," says de Graf, "is not going to be interesting unless it is populated with interesting characters."

Pattie Maes, an artificial life researcher at the MIT Media Lab, abhors goggle-and-gloves virtual reality. She finds such clothing "too artificial" and confining. She and colleague Sandy Pentland came up with an alternative way to interact with virtual creatures. Her system, called ALIVE, lets a human play with animated creatures via a computer screen and video camera. The camera points back at the human participant, inserting the observer into the virtual world that he or she is watching on the screen.

This neat trick gives a real sense of intimacy. By moving my arms I can interact with little "hamsters" on the screen. The hamsters look like tiny toasters on wheels, but they are autonomous goal-seeking animats that contain a rich repertoire of motivations, sensors, and responses. The hamsters roam the enclosed pen looking for "food" when they haven't eaten in a while. They seek each other's company; sometimes they chase each other. They run from my hand if I move it too fast. If I move it slowly, they try to follow it out of curiosity. A hamster will sit up and beg for food. When they get tired, they fall over and sleep. They are halfway between robots and animated animals, and only several steps away from authentic virtual characters.

Pattie Maes is trying to teach creatures "how to do the right thing." She wants her creatures to learn from their experiences in the environment, without much human supervision. The Jurassic dinosaurs won't be real characters until they can learn. It will be hardly worth creating a humanist virtual actor unless he or she could learn. Following the subsumption architecture model, Maes is structuring a hierarchy of algorithms that let her creatures not only adapt, but also bootstrap themselves to increasing complex behaviors and -- as an essential part of the package -- also let their own goals emerge from those behaviors

The animators at Disney and Pixar nearly croak at the thought, but someday Mickey Mouse will have his own agenda.