Let there be light! Robot machine vision leapsforward thanks to real timeimage processing.

Ten years ago, processor speed was not fastenough to enable real time video imageprocessing. Today, it is. “Real time processing ofvideo images means we can now put cameras onrobots and have them understand the worldthrough video images,” says Professor Peter Stone,Department of Computer Sciences, University ofTexas at Austin.

A decade or more ago, roboticists could onlyuse sonar range finders in a single dimension todiscover how far the robot was from the nearestwall. This gave the robots no information aboutfaces for face recognition, no information about

Contact the author at geercom@windstream.netby David Geer

Artificial Intelligence BringsHumanoid Robots to Life

Researchers have a variety of goals for artificial intelligence. Some scientists are examining human biology together with artificial intelligence in an effort to create robots that are as humanlike as possible. Still others examine AI from an engineering perspective, begging the question,

what can robots be enabled to do?

The mere building blocks for advancing AI for humanoid robots are many andhighly technical. Robotic vision via webcams and machine vision, naturallanguage processing and machinelearning, common sense reasoning andknowledge, robotic agility and motion,and robot-to-human-to-robot interactionare a few of the areas of research that will move us closer to more human-likerobots. What research areas will have toemerge to take us there? Will we knowbefore we get there?

12 SERVO 01.2010

Nao robot shown withhuman hand (for scale).

Geerhead - Jan10.qxd 12/6/2009 5:08 PM Page 12

people’s location relative to the robot, and noinformation about color.

Robots moving from legacy sensors to realtime machine vision is like asking people to walkaround a room with their eyes closed, illustratesProfessor Stone. People would likely use theirhands as bump sensors to sense objects and feeltheir way around the room.

This would make moving around slow,complex, and herky-jerky. Humans can do a lotmore if they are allowed to open their eyes andprocess vision in real time. So can robots.

Robot machine vision has made a lot ofprogress, though not all the problems surroundingit have been resolved. At the same time, there hasbeen progress accomplishing vision tasks using anew three-dimensional laser range finder thatretrieves more than 1M range points per second.This gives the robot 3D range pictures of theworld around it, according to Professor Stone.“We are researching this here,” he comments.

Velodyne is making this range finder — theLIDAR HDL-64E. The 64-element sensor offers 360-degrees of HFOV (horizontal field of view) and26.8 degrees of VFOV (vertical field of view),according to Velodyne. It reads 1.3M data pointsper second.

This provides the distance and intensity dataabout the environment that are necessary toachieve a 3D view of the world.

While the HDL-64E is too big for a smallhumanoid robot, it might fit onto a full-sizedhumanoid robot, according to Professor Stone.“We have an HDL-64E on one of our autonomouscars. We are able to get a much richer stream ofinformation than was ever possible before,” hecontinues.

Learn the LanguageFor humanoid robots to become as intelligent

as people, they will have to be imbued with theability to understand natural human language. Thisis called natural language processing. “Right now,if you type a question into Google search, it won’tsimply answer your question. It doesn’tunderstand language in the way that people do,”explains Professor Stone.

With robots, the idea would be that theycould hear or read natural human language,understand it, and respond with learning,communication, or obedience to a command.With computers, people could stop usingkeyboards and simply talk to their computers. Thecomputers would be able to listen, parse the

words and sentences, and understand andrespond by typing out words in a word processoror initiating computer commands.

“You could just tell your computer, open mybrowser and search for directions to the nearestrestaurant, rather than having to go and do whatwe do in Google maps,” says Professor Stone.

Machine LearningWhen robots learn by doing, they no longer

need to be instructed or controlled quite so muchby an outside intelligence (man’s). “With machinelearning — which is one of my areas of expertise —the goal is to write and develop algorithms that

SERVO 01.2010 13

GEERHEAD

Here’s lookin’ at you, Nao.

Professor Peter Stone of the Department of Computer Sciences,the University of Texas at Austin, is dedicated to the goal of theRoboCup Federation, which is to build a humanoid robot team thatcan defeat the best human world cup team on a real soccer field by2050. According to Dan Burrus, founder of Burrus ResearchAssociates, Inc., and a long-time roboticist, it is more likely thatProfessor Stone will reach his goal by the 2030 to 2040 time frame.

Accomplishing Professor Stone’s goal will solve many importantchallenges and problems around AI for humanoid robotics. Examplesinclude the robot agility problem (having robots run quickly andmanipulate their motors in a very granular way so that they can kick a soccer ball), planning, teamwork, and higher level cognitivefunctioning. “The robots need to be able to reason about theiropponents and not fall for the same trick over and over,” ProfessorStone explains.

News: Humanoids Beat Humans in 2050

Geerhead - Jan10.qxd 12/6/2009 5:09 PM Page 13

allow computers and robots to improve theirperformance over time without being explicitly toldhow to accomplish their tasks,” says ProfessorStone. Rather than Professor Stone and hisstudents programming their robots to walk faster,to move their legs a certain way, to move intocertain joint angles, they programmed in amachine learning algorithm that gave them theability to experiment with walking; to try differentleg motions and joint angles. With this, the robotscan record the movements that enable them towalk faster. They can then experiment withcombinations of these movements, to see which ofthose help them to walk faster. In this way, therobots can improve the quality of their behavior inrelation to the desired goal.

Processors, Sensors, and Bandwidth

The more powerful the building blocks — suchas processing power and speed, sensor capability,and bandwidth — the more scientists are able tothink in new and different ways about solvingproblems about AI. Here, Professor Stone draws ananalogy with man’s attempt to use computers toplay chess:

“When computers were first developed, thenotion of a computer that had the speed of DeepBlue really wasn’t imaginable. But, at the time,people were trying to figure out whethercomputers could be made to play chess using thetypes of processing speeds that people had backthen, in the 1970s, which is just orders and ordersof magnitude different from what is availabletoday.”

“Once you get to the point of being able todo the kinds of search operations that computerscan today, then a more or less brute force type ofapproach, searching billions of chess boardpositions every second, is made possible, and itopened up completely different ways of thinkingabout the problem,” says Professor Stone.

AI TodayToday, artificial intelligence is good at all the

things that modern day computers are good at,such as searching through databases ofinformation looking for the most relevant records.“Humanoid robots really are computers,” saysProfessor Stone. Robots are not better than peopleat physical motion, though that is what weprimarily think of robots as doing.

Robots do not yet have the same degrees offreedom that humans have. Researchers areworking on this, including robots with degrees offreedom in their backs, through use of spines thatmimic the capabilities of human spines, accordingto Professor Stone. “But, giving robots spines isdifficult and expensive.”

ConclusionIn order for truly humanlike humanoid robots

to model the capabilities of people, they must beinstilled with intelligence for a number of individualtasks and accomplishments. Taken together,however, these different types of intelligence willform building blocks for more fully humanlikebehaviors and thinking. SV

GEERHEAD

14 SERVO 01.2010

Front view of a Nao robot, like the ones Professor Stone and others are experimenting with for RoboCup soccer.

Video of Nao robots used in RoboCupwww.cs.utexas.edu/~AustinVilla/?p=nao

Geerhead - Jan10.qxd 12/6/2009 5:09 PM Page 14