Pittsburgh’s Leadership in Artificial Intelligence Is Spawning the Technologies of Tomorrow

“JACK” – The Seegrid Intelligent Pallet Jack maps the ceiling as it traverses a previously traveled path and compares it to images stored in its map atlas to find its way around a warehouse.

by Tom Imerito

As I passed through the double doors to Seegrid’s manufacturing floor a driverless robotic pallet jack that was headed toward me stopped abruptly.  A horizontal bank of orange LEDs blinking above a row of steadily lit green ones at the top of its control module gave me the impression that the eight-foot mechanical giant was thinking – which it was – but not about me.   My host, Seegrid’s Director of Research and Development of Controls and Capabilities, Dr. Darrin Bentivegna, told me that the machine was going through a pre-shipment burn-in.  I had coincidentally entered the room just as “Jack,” (as I came to anthropomorphize the robotic Titan) was pausing in response to a command in his memory

I was excited, because I was about to train Jack to travel a new course.  It would be my first hands-on experience with an intelligent machine.  Training Jack was as easy as grasping his control handle and leading him around a corner, stopping to drop of a pallet, sounding his horn and returning to the starting point.

As I walked ahead of him, the four pairs of stereographic cameras mounted beneath his LEDs captured hundreds of images of the surrounding environment.  As we walked, Jack’s onboard processor converted the images into digital maps which he would retrieve from his memory and compare with the images he collected on his next trip along the newly navigated course.

When we finished the training run I sent Jack off on his own.  He retraced the course perfectly, gathering new image maps as he went and comparing them with the ones he had saved during our training run.  In response to differences in the maps he continuously corrected his course while creating new maps of his surroundings for use on his next trip.

Jack’s comparison of the maps in his memory to the maps he sees at any given moment and his consequent course-correction is emblematic of artificial intelligence (AI).  At its most basic level AI compares data from different sources and modifies activity based upon the difference between them.  Data can be created on the fly by sensing the environment, as with jack’s stereo cameras, or it can be searched and retrieved from a data base, as with jack’s memory atlas.

Although the differences between the way machines and humans think are vast, it is convenient to think about artificial intelligence as machines learning in a step-wise fashion by – 1) acquiring data by either sensing or searching; 2) comparing the data to find the most likely solution to a problem; 3) deciding on the best course of action (including providing an answer, changing the formula or algorithm being used to solve the problem, looking somewhere else or starting all over and trying something else). Although such a tedious way of thinking about everything we do would likely drive most human minds crazy, computers are so much faster than we humans they can do tedious tasks over and over again without a blink of their a digital-camera eyes.

Jack is the embodiment of the vision of Seegrid founder and world renowned roboticist, Hans Moravec who, in his 1988 book, Mind Children, envisioned a future world in which intelligent machines perform the mundane tasks of civilized life.  Today Moravec’s vision is becoming reality in Pittsburgh where, in addition to Seegrid’s’ robotic industrial trucks, a host of artificial intelligence companies are engaged in activities as varied as building instant online communities, translating languages, training health care professionals, teaching math, improving driver performance, inspecting sewers and helping people lose weight and stay fit.


CMU spinout, FlashGroup creates instantaneous topical interest communities based upon a user’s current browsing habits.  The company’s widget scours a user’s recent web history for the central topic of interest while he or she is engaged with it then searches the web for related resources.  By comparing a user’s current topic with millions of related resources such as expert web pages, social media discussions and individuals looking at similar web pages at the same moment, FlashGroup provides an instantaneous battery of resources related to a user’s interest.  Just as Jack compares maps, FlashGroup compares user behavior with related web resources.  Unlike search engines which rely on words typed by a user, FlashGroup’s widget decides what a user is interested in by assessing web activity on a moment-by-moment basis.

The company is currently working on a confidential basis with a number of well known companies and plans to open its digital gateway to smaller users soon.


Taking AI on a tour of the world, Carnegie Mellon spinoff Safaba Translation Solutions employs parallel text processing to automatically translate corporate communications from one language to another at levels of accuracy approaching human translations.   Similar to the way Jack compares one map to another to navigate, Safaba compares large bodies of human-translated digital texts in both the original and translated versions to produce a customized translation engine.  The translation engine utilizes statistical models that it learns from the side-by-side comparison of texts in two languages to translate new texts automatically.  Safaba fills a niche in corporate localization efforts by enabling the high-speed translation of institutional communications such as web pages, chat sessions and other real-time communications into a unique a corporate language which is highly precise at the same time it is friendly and comfortable to local stakeholders around the world.

The company is currently providing automated translation services to several household-name global organizations.


Employing a form of artificial intelligence called synthetic interview, CMU spinout, MedRespond has developed an inexpensive training application for health care professionals.  The MedRespond system is composed of a data base of hundreds of pre-recorded video responses to thousands of issues typically posed by patients to health care professionals such as doctors, pharmacists, nurses, and technicians.

In practice, an actor/patient presents questions, symptoms and responses in the form of pre-recorded videos, to which a health care provider responds by keyboard in real time.  The response is analyzed by another AI technique, called natural language processing which analyzes the words in a sentence by occurrence, meaning, position, frequency, and proximity to other words, to establish both the literal meaning of the response, as well as its tone.  As a result of its analysis, the program makes a decision about which video response to present next.  The machine’s response leads to a reply by the human trainee, which, in turn leads to another – just like a human conversation.  MedRespond vastly reduces the cost of training health care professionals by eliminating the redundancy of human trainers responding over and over to the same issues, face-to-face, with hundreds or thousands of health care professionals.

The company is currently conducting a pharmacist training program with drug retailer Rite Aid Corporation.


If AI can be used to train health care professionals, it would seem a simple matter to use it to teach kids in school.  Located in downtown Pittsburgh, Think Though Math (TTM) is doing just that.  The company is web-tutoring students in math in grades four through first-year algebra.  The TTM system uses online tests to determine a student’s performance level or zone of proximal development (ZPD) and gauges lessons and tests to make the subject matter ever-so-slightly beyond a student’s current grasp, but definitely reachable with some effort.  TTM’s algorithm compares an individual student’s accuracy and response time with normalized tables to determine when a he or she needs a hint, a chapter review or, when completely stumped, intervention by a human tutor.  Upon meeting the algorithm’s standard for subject mastery the student goes to the next level.  Students earn points redeemable for consumer products as a reward for excellence.

According to the company’s vice president of marketing, Peter Cipkowski, “Think Through Math is having a positive impact on tens of thousands of students every day – and growing.”


If earning points can get school students to perform better, shouldn’t it work for – let’s say truck drivers?  Pittsburgh’s Propel IT is proving that it most cases it does.  Propel IT employs the data already captured by the on-board computers legally required in every over-the-road truck to increase fuel efficiency by modifying driver behavior.  By collecting about 100 of the 4,000 variables measured by a truck’s computers, and comparing them with a self-learning data base of driving behaviors and fuel efficiencies, the Propel IT system decides who is driving efficiently and who is not.  The system’s machine-learning algorithm modifies itself based upon differences between a driver’s behavior, the nature of the load and trip as well as fleet-wide performance.   Off the road, the system suggests ways for drivers to improve their personal driving habits and allows them to earn points redeemable for cash and gifts for improved performance.

Of the 25, 000 trucks currently using or testing Propel IT, fuel consumption has dropped by an average of five percent.  When multiplied out over the 8 million trucks on U.S. highways each running 100,000 miles a year, Propel IT’s potential national energy savings comes to more than $37 billion in diesel fuel per year.


Easily the least glamorous, but possibly the most profitable of Pittsburgh’s AI ventures, RedZone Robotics provides services that take place beneath the road – robotic sewer inspection services.  RedZone’s robots detect the condition of a sewer pipe’s structural integrity and interior condition with digital cameras, laser scanners, chemical and inertial sensors, and sonar to create a digital record of a length of pipe ranging from a few feet to a several miles.  The autonomous devices are powered by lithium ion batteries, controlled by software in an on-board CPU, and propelled by neoprene tracks.

Once lowered into a sewer the machine follows a pipe for a pre-programmed distance, gathering visual, chemical and mechanical information about the pipe’s geographic location, physical condition, damage and obstructions – information which is virtually impossible to know short of digging it up after a catastrophic failure.  At the end of its pre-programmed trip, the completely wireless robot returns to the surface, where it automatically uploads its data to the cloud for integration with existing municipal infrastructure data to form a comprehensive overview of an entire sewer system.  RedZone’s resulting maps allow city managers to view an interactive street map of a municipality and click anywhere to view a video and quantified data about the condition of the sewer system beneath the street.

According to RedZone CEO, Eric Close, the company currently serves 250 municipalities around the world.   With an estimated annual sewer renovation market of $11.5 billion and 740,000 miles of sewer lines in the United States, most of which have never been visually inspected, RedZone expects to see continued enormous growth for the company.


Because large amounts of data tend to ensure greater accuracy of machine-generated results, in AI the rule is – the more data the better.  With a database of human energy intake and expenditure comprising millions of user-days of activity, BodyMedia’s Free-Living Energy Expenditure Database is believed to be the largest such aggregation of data of its type.

The company’s arm-worn multi-sensor energy monitor detects body heat, perspiration and motion at a rate of thousands of data points per minute to determine the number of calories a user is burning. Data is uploaded to a user’s Activity Manager – the company’s web-based, online tool – and compared with his or her food log to determine a net calorie surplus or deficit, from which eating habits and exercise regimes can be maintained or modified.

In 2011 the company was awarded the Innovative Applications award by the International Association for the Advancement of Artificial Intelligence.   BodyMedia Armbands are available online and at a many retail stores, including Giant Eagle, Target, Amazon.com and Best Buy.


If two decades ago, Hans Moravec’s vision for the future looked like pie-in-the-sky speculation, today artificial intelligence looks like a very real industry – especially in Pittsburgh.


This article first appeared in the Pittsburgh Technology Council’s TechBurgher Magazine.