Page 16 - ME-06-E

ing multi-objective genetic algorithms and reinforce- time at Duke University, she discovered that neuronal

ment learning. Finally, we demonstrated the resulting activity is higher in the early sessions, when the inter-

walking ability on robots we built at our lab using a face is first used. "We developed a theory that links

Research locomotion mechanism that utilizes the leg's natural this increase in neuronal activity with the brain's

dynamics. response to mistakes the interface makes because

Today, Prof. Zacksenhouse's lab focuses on rein- of prediction errors. Our goal is to identify the brain

forcement learning for controls in robots performing activity that happens in response to a mistake, and

assembly tasks. "For robots to be economically fea- train the interface to correct those mistakes and im-

sible, they have to be able to handle uncertainties in prove its performance."

part locations and learn to generalize from one action Today, Zacksenhouse's lab develops non-inva-

to another. These requirements are even more im- sive brain-machine interfaces that use EEG measure-

portant for small and medium-sized factories, where ments from electrodes placed on the scalp. One such

the small amount of products of each type makes it project is a brain-machine interface for bending and

unfeasible to prepare an exact production line and straightening prosthetic arms. "The interface classi-

program the robots manually. Using reinforcement fies the EEG signals to decide whether to change the

learning, we teach robots to perform tasks like hard- position of each arm. We use error-related potentials

ness control, train them to independently learn the – signals that occur when the brain responds to the

parameters for different tasks, generalize what they'd machine making an error – to improve performance.

learned, apply it to new tasks, and handle uncertain- In effect, this brain activity gives feedback to the ma-

ties. For this project, we are collaborating with com- chine, thus allowing it to correct its mistakes and in-

panies in the industry and receiving funding from the crease its accuracy."

Israeli Innovation Authority." Zacksenhouse and other researchers at her lab are

In her second lab, Prof. Zacksenhouse studies also collaborating with a lab in Michigan, helping im-

brain-machine interfaces – direct communication prove the invasive brain-machine interfaces that are

being developed there. Funded by

The D. Dan and Betty Kahn Founda-

Prof. Miriam Zacksenhouse: tion for Michigan-Israel Collabora-
"For robots to be economically tions, the study aims to improve the
feasible, they have to be able performance of invasive brain-ma-
to handle uncertainty. Using chine interfaces used to help peo-
reinforcement learning, we ple with locked-in syndrome and
train them to teach themselves to power exoskeletons, which help
the parameters for a specific able-bodied people do physically
task, generalize what they'd demanding work.
learned, apply it to other tasks,
and handle uncertainty" The research students at Zack-
senhouse's lab have very diverse

academic backgrounds: "I have

students who studied mechanical

engineering, electrical engineer-

ing, aeronautical engineering, bio-

medical engineering, and medical

science," she says. Some of our

alumni have started their own com-

channels between the brain and the outside world. panies. One such researcher is Asst. Prof Jonathan

"We began our work in this area after my sabbati- Spitz, whose startup uses machine learning to devel-

cal in 2003 – a year I spent at Duke University, at a op industrial robots. Others work for companies that

lab studying invasive brain-machine interfaces that develop invasive brain-machine interfaces (like Elon

worked by measuring the activity of hundreds of sin- Musk's Neuralink) and design algorithms that har-

gle neurons using electrodes implanted in the brain," ness the power of machine learning for the robotics

she tells us. These interfaces are used for operating industry at global giants like Bosch as well as our own

advanced prostheses, like bionic arms. During her local leaders, like Elbit.

16 | MEgazine | Faculty of Mechanical Engineering