Beyond mimicry : Transfer on robots the human concepts of learning.

How to make I.A. even more humain ?

The company Vicarious has launched a research project in the field of robotics science that will endow the artificial intelligence of an almost human brain. Read the article to the end to find out more about Vicarious. This project is being implemented by researchers Miguel Làzaro-Gredilla*, Dianhuan Lin, J. Swaroop Guntupalli and Dileep George*. It was published on January 16, 2019.

image from the site – Vicarious

The aim of this project is to be able to improve the robots’ ability to understand our intent and to transfer tasks between different environnments. In order to achieve this goal, researchers have designed a program capable of replicating aspects of human concept learning. Here, concepts are to be seen as programs that are put into a computer whose architecture consists of a system of visual perception, memory work and action controller. The team of researchers demonstrates how a robot can interpret new concepts from abstractions and how it can apply them in defferent situations.

These scientists have managed to get closer to the goal of building robots with a strong ability to understand. This was possible by bringing to their project the ideas of cognitive science and neuroscience.

Their hypothesize are that human concepts are programs and computers can acquire the same characteristics as the human brain. In this direction, the team of scientists developed a computer architecture called visual cognitive computer (VCC) and represented concepts as programs on this computer.

A neuroscience-based approach

The researchers started by inducing concepts into the computer. They gave to the VCC two images that represent a concept and then the computer must understand this concept to run it in the physic world. Once he has understood the concept, the computer is capable, only by viewing the first image, to predict « imagine » what will be the outcome, the second image.

To illustrate this experience, you can use the LEGO or IKEA furniture assembly manuals as an exemple. It’s a sequence of instructions and concepts of logic which the end result can easily be imagined, except for computers who show more difficulty in understanding than humans ; well, for now.

image from the file – Science Journals — AAAS

Then, they have done a series of complex tests on the program to succeed in improving it.

At the end, they have transferred the program into two different robots : Baxter from Rethink Robotics and a UR5 from Universal Robot, and tested the transfer of five different induced concepts on the two robots, and each concept was tested in several different scenarios.

Almost perfect results

The robot UR5 succeeded in executing the concepts in more than 90% of the trials. The failures were caused by the objects slipping out of the gripper during a grasp or during placement. On Baxter, the success rate was lower (at about 70%) due to a blurry camera, a less effective gripper, and imprecise motion in our aging robot.

So, scientists have managed to make a cognitive program based on the ideas of neurosciences and tested it on 546 different concepts, 535 of them could be learned with a search budget of 1 million programs, and 526 whith a search budget of 0,5 million programs.

Their experiments show that with the program they designed, it is possible to make a robot understand the task to execute in the real world without any explicit demonstration.

the goals in artificial intelligence and robotics

« Getting robots to perform tasks without explicit programming is one of the goals in artificial intelligence and robotics » explained the researchers. This goal is now achieved through this resarch program.

To sum up, they have succeeded in developing a program that can reflect and learn by using the computational principles of the human brain. It’s algorithmic architecture focuses on the disciplines of vision, langage and motor skills.

At the end of 2013, the software was able to solve more than 90 % of the Captcha (anti-spam visual system that requires copying a distorted string) while it has not been programmed in this sense.

Through this kind of research project, robotics and artificial intelligence are constantly advancing. We are now at a stage where robots are capable of more than just imitating the human brain.

Some information about the Vicarious company

Vicarious is an innovative company created in 2010 based in San Francisco, California. It has 23 employees and 70 millions dollars.

It received investments from Mark Zuckerberg, Jeff Bezos, Elon Musk, Peter Thiel, Sam Altman, Samsung and ABB Robotics.

To remember

  • Researchers from Vicarious have developed a program capable of more than simulating the human brain.
  • Artificial Intelligence and robots can now understand our intentions easily.

Links and sources :

https://www.vicarious.com/wp-content/uploads/2019/01/10.1126@scirobotics.aav3150.pdf

0 thoughts on “Beyond mimicry : Transfer on robots the human concepts of learning.”

Leave a Reply

Your email address will not be published. Required fields are marked *