Thanks to Deep Learning, this Robot can Perform Tasks with Almost Human Level Skills
Overview
- Researchers at UC Berkeley have created a robot that can sift through objects and pick them with amazing accuracy
- Called, Dex-Net, the software has been developed with deep learning through a deep neural network
- A new metric, called mean picks per hour, is being used to measure and improve it’s performance
Introduction
AI powered robots have been around for a while. We’ve covered robots that can cook the perfect burger, read the news, and even deliver speeches. But this latest creation, by a professor and his student from UC Berkeley, is the smartest one developed so far.
It has a range of functions, but the most useful one is that it can sort through your junk drawer and pick out stuff with the speed and accuracy that rivals humans. The key to it’s smartness lies in how the software behind it has been programmed.
Source: MIT Technology Review
The software is called ‘Dex-Net’ and it identifies how to differentiate and pick-up objects, especially ones that might not belong in the pile, with unerring accuracy. You might be wondering at this point how does the robot actually work?
Well, this is where deep learning comes into play. The software powering the robot has been developed with a deep neural network – it designed to pick up objects in a virtual environment, sort through them, and continuously learn with each picking. But here is what separates it from anything done previously in this field – Dex-Net generalises from an object it has seen before to a new one. What makes it even more human is that if it’s not able to understand what the object it, the robot will look at it from another angle to get a better view of it.
As you can see in the image above, the robot has two arms that are powered by a different neural network. One arm is supposed to grasp objects while the other arm performs the function of suction.
The two researchers also came up with a way to calculate how to measure the performance of Dex-Net – using a measure called “mean picks per hour”. Basically, this is calculated by multiplying the mean time per pick and the mean probability of success. Humans are capable of 400-600 mean picks per hour while this robot has reached 200-300 so far.
A demonstration of how the robot works was given at MIT Technology Review’s EmTech Digital event. You can view a video of the robot strutting it’s stuff below:
Information from MIT Technology Review’s blog post was taken in this article.
Our take on this
This robot is as close to matching human skills and understanding as anything created before. Once it has been programmed with even better commands, it can potentially be used for diverse functions, both commercially and in the industry.
And here’s the advantage it has over humans – Dex-Net does not get tired and can relentlessly plough through objects without experiencing any lag time. With the mean picks per hour metric, it has given researchers something to measure and improve with each iteration of the software.
In addition to Berkeley’s efforts, researchers at DeepMind and OpenAI are also experimetning with robotic technology and new and even mroe intelligent AI is expected to be revealed soon.
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!
Also, go ahead and participate in our Hackathons, including the DataHack Premier League and Lord of the Machines!
Senior Editor at Analytics Vidhya. Data visualization practitioner who loves reading and delving deeper into the data science and machine learning arts. Always looking for new ways to improve processes using ML and AI.
