Nouvelle AI

• 1. Introduction

  ›

  Nouvelle AI is an approach to artificial intelligence that Rodney Brooks pioneered in the 1980s at MIT's artificial intelligence laboratory. At a time when most researchers were building AI systems crammed full of symbolic descriptions and elaborate internal maps, Brooks asked a different question: what if the world itself could be the model? What if a machine did not need to store a picture of its environment at all, but simply sensed and reacted? And what if that was enough to produce genuine intelligence?

  The robots that came before Brooks had a telling problem. Shakey, one of the earliest symbolic AI robots, needed to run planning programs that broke every desired action into individual steps. The computation required so much time that Shakey moved through the world painfully slowly. Its internal model of the micro-world had to be updated constantly as the robot moved or as objects shifted around it.

  Brooks believed intelligence could emerge organically from simple behaviors as those behaviors interacted with the real world. That idea raised a question the rest of this documentary will explore: can something as modest as an insect-level mind point the way toward human intelligence?

• 2. The Symbolic AI Problem

  ›

  Shakey and Freddy, two early robots built on symbolic AI principles, each carried an internal model of their environment built from symbolic descriptions. Every change in the robot's position, or any shift in the surrounding world, forced the system to renew that structure of symbols. The overhead was relentless.

  Symbolic AI researchers had long struggled with updating, searching, and manipulating these internal worlds. The technical name for one version of this burden is the frame problem. Representing a robot's state using first-order logic requires a large number of axioms just to convey that things in an environment do not arbitrarily change. The logical machinery needed to handle even simple facts about stability ballooned quickly.

  Brooks offered a different path. A nouvelle system refers continuously to its sensors rather than to an internal model. As Brooks put it directly, "the world is its own best model--always exactly up to date and complete in every detail." By letting the real world carry the information load, a nouvelle robot sidesteps the need to fill the machine with volumes of symbolic language.

• 3. Simple Behaviors, Complex Results

  ›

  The central idea behind nouvelle AI is that simple behaviors combine over time to produce more complex ones. Take two basic instructions: move forward, and avoid obstacles. Pair those with a tendency to move toward a moving object, and something that looks like chasing can emerge without ever being explicitly programmed.

  Herbert, one of Brooks's insectoid robots named after cognitive science and AI pioneer Herbert A. Simon, made this principle concrete. Herbert used infrared sensors to avoid obstacles and a laser system to collect three-dimensional data over a distance of about 12 feet. It also carried simple sensors in its hand. Its testing ground was the busy offices and workspaces of the MIT AI lab, where it searched for empty soda cans and carried them away.

  That soda-can behavior emerged entirely from 15 simple behavior units combining. Herbert never stored information for more than two seconds and discarded a high volume of what its sensors picked up. Simon himself noted a parallel: an ant's complicated path through terrain reflects the structure of its environment, not the depth of its own thought processes.

• 4. Allen, Genghis, and Squirt

  ›

  Allen was named after cognitive science pioneer Allen Newell. It carried a ring of twelve ultrasonic sonars as its primary sensors and three independent behavior-producing modules, all programmed to avoid stationary and moving objects. With only that single module active, Allen would stay in the center of a room until an object approached, then move away while steering around any obstacles in its path.

  Genghis took the insectoid design further. With six legs, it was able to walk over rough terrain and track a human. Squirt operated on a simpler set of triggers: its behavior modules kept it hiding in dark corners until a noise occurred, at which point it would move toward the source.

  Brooks eventually agreed that nouvelle AI had come close to the complexity of a real insect. That concession immediately raised a harder question: was insect-level behavior a reasonable stopping point, or merely a starting line?

• 5. Embodiment and Turing

  ›

  Traditional AI had aimed to build intelligences without physical bodies, systems that could only interact with the world through a keyboard, screen, or printer. Brooks pushed in the opposite direction. Nouvelle AI attempts to build what researchers call embodied intelligence, a mind situated in and responding to the real world.

  Brooks found support for this approach in Alan Turing's own early sketches. In writings from 1948 and 1950, Turing described a "situated" approach that involved equipping a machine "with the best sense organs that money can buy" and teaching it "to understand and speak English" through a process that would "follow the normal teaching of a child." Turing contrasted this with approaches focused on abstract activities such as chess.

  That contrast carries a pointed implication. Chess-playing AI operates in a closed, well-defined symbolic world. A machine learning language and perception the way a child does is engaging with genuine open-ended complexity, the kind of complexity nouvelle AI was designed to embrace.

• 6. Cog and the Human Question

  ›

  In the 1990s, Brooks shifted his ambitions upward. Together with Lynn Andrea Stein, he built a humanoid robot called Cog. Cog had an extensive collection of sensors, a face, and arms, features designed to let it interact with the world and accumulate experience in the organic way Turing had described.

  The decision put Brooks in an interesting position relative to John Von Neumann, who had argued that theorists who chose the human nervous system as their model were unrealistically picking "the most complicated object under the sun." Von Neumann also dismissed the ant as a model, reasoning that any nervous system exhibits exceptional complexity. Brooks nonetheless pursued Cog, betting that situated, embodied learning could build common-sense knowledge from the ground up.

  The Cog project stopped receiving development in 2003. Whether the robot ever found meaningful correlations between its sensory inputs and its actions, and what those results suggested about machine learning through embodied experience, remained open questions when the work came to a halt.