12_5_14_Hadas Kress-Gazit_0097.jpg

Graduate student Jonathan DeCastro and Robot Professor Hadas Kress-Gazit and her research team are working on it.
Frank Dimeo
Frank Dimeo

12_5_14_Hadas Kress-Gazit_0082.jpg

Kress-Gazit’s team discusses programming reliable robots. Kress-Gazit’s lab tackles programming reliable robots. “People write code, they test it, and at some point they’re confident…it’s ready…But if there’s a bug in the system…they will never know until the robot fails.”
Frank DiMeo
Frank DiMeo

12_5_14_Hadas Kress-Gazit_0014.jpg

Hadas Kress-Gazit The question that fascinates me is how can we control big, complicated robots and make them do a variety of tasks in a safe and predictable way.
Frank DiMeo
Frank DiMeo

12_5_14_Hadas Kress-Gazit_0064.jpg

Robots see no evil, hear no evil, speak no evil. Robots see no evil, hear no evil, speak no evil.
Frank DiMeo
Frank DiMeo

12_5_14_Hadas Kress-Gazit_0036.jpg

The Kress-Gazit lab One of the goals of the Kress-Gazit lab is “Rather than having a person programming the machine for months, you just tell it what you want it to do”…like JARVIS.
Frank DiMeo
Frank DiMeo

Ready for Your Very Own Robot?

by Lauren Cahoon Roberts

Anyone who’s watched Iron Man is probably familiar with JARVIS, Tony Stark’s right-hand artificial-intelligence entity. The robot is as witty and urbane as a British butler, assisting, responding—even reprimanding Stark when the occasion calls for it and responding to any commands with unquestioned intelligence. If Hadas Kress-Gazit, Mechanical and Aerospace Engineering, had her way, JARVIS would go from a Marvel Comics creation to a tangible technology that you or I could access. “Right now, robotics is still a niche. The robots we have can only do specific things,” says Kress-Gazit. “The question that fascinates me the most is how can we control these kinds of big, complicated robots and make them do a variety of tasks in a safe and predictable way.”

Instructing Robots, Naturally

As Kress-Gazit points out, we currently have robots, such as the Roomba, that can do one simple thing well (vacuum a room) and do it autonomously. On the other side of the spectrum, we have robots that can perform complicated tasks under close human supervision. These robots have been programmed by experts and cannot deviate from that program. Interacting with robots on either end of the spectrum requires specific programming or instructions, rather than an informal, verbal request. These issues present the two most pressing questions in Kress-Gazit’s research: How can we create robots that can be programmed by non-experts through natural, spoken language, and how can we ensure they behave correctly?

Kress-Gazit is hard at work answering these questions. This field of study has been a natural path for her. As a young adult, she was an avid sci-fi and fantasy reader, so naturally, “I found robots to be pretty cool,” she shares. She developed a more intellectual excitement for robotics as she studied electrical engineering as an undergraduate at Technion (Israel Institute of Technology). After working in chip design for IBM and missile design for a defense contractor, Kress-Gazit decided to pursue robotics academically as an MS and PhD student at the University of Pennsylvania. “It was just always a cool subject to me. I liked the multidisciplinary aspect to the field of robotics.”

Reliable Robots

Her interest in addressing robot programming begins at the root of a fundamental issue with robotics today. “Currently, people write code, they test it, and at some point they’re confident that it’s ready,” she says. “But if there’s a bug in the system, and something could throw that robot off, they will never know until the robot fails.” Kress-Gazit wants to change this paradigm by developing new algorithms that can guarantee success for a high-level task that’s been requested by a human. “My research won’t say, ‘okay, your robot is perfect,’ but it will provide assurances. It will show that it’s safe, and where it could fail.” To do this, her team uses synthesis and verification techniques, including logic, automata theory, and control theory to transform high-level tasks into robot controllers.

How can we create robots that can be programmed by non-experts through natural, spoken language, and how can we ensure they behave correctly?

What’s more, Kress-Gazit would like to ensure that her robots do not require the kind of nitty-gritty programming expertise that most machines do today. “The big picture of this idea is to allow anyone to be able to program robots—so that my nine-year-old child could tell a robot to do a task and have it respond correctly. And rather than having a person programming the machine for months, you just tell it what you want it to do,” she says—like JARVIS. “We want to create a robot controller that is guaranteed to do what it’s asked to do in an intelligent way."

Kress-Gazit’s team uses techniques from computational linguistics to map language to a logical formalism known as Linear Temporal Logic, which enables them to synthesize controllers. These techniques allow them to provide specific feedback when things go wrong and to explain what the robot is trying to achieve at any given time. Ultimately, all the algorithms, coding, and testing come down to one driving aspiration for Kress-Gazit: “I’d like to enable the creation of a robot that has impact on society in a meaningful way.”

Anyone nervous about Kress-Gazit and her colleagues’ work leading to a dystopian rise of the machines should rest easy; she is not concerned. “Despite the sci-fi appeal, getting robots to do simple things like turning a valve or connecting a hose is incredibly difficult today, as the DARPA Robotics Challenge is showing us,” she says. “Yes, we will make great technological advances in the coming years, but we will also make progress on how to guarantee these machines will remain safe and predictable.”