Robots play an irreplaceable role in modern manufacturing. However, safer, cheaper, and especially smarter robots are in short supply.
With the rapid development of materials science, computer science, brain science and other fields, although robots are far from the image in science fiction movies, with the continuous improvement of "learning" ability, the development and expansion of robots and the real world interactive technology Practice, but it shocks you and my eyes all the time. Although each study has only improved a little, it is this "a little bit" progress that will eventually converge into a smart world of the future.
The US MIT Technology Review has published several articles on new trends in robot learning capabilities since October, which may be an interesting reference for robot enthusiasts to track the latest technology directions.
"Self-learning" to grab a variety of items
One of the goals of general-purpose robots is to interact intelligently with the objects in everyday life, but the robot's ability to grab is really embarrassing. Let the robot pick up a TV remote, a bottle of water, or a toy gun, and it will continually explore it unless it has a program that allows it to pick up specific items in a particular environment.
This is in stark contrast to the ability of humans to grab things. A human baby can quickly learn to grab specific items in the most chaotic and unstructured environment.
So, can robots learn to grasp by trial and error and error correction like a baby?
Today, Carrefour Mellon University's Rirel Pinto and Aposinavi Gunta Pita prove this possibility. They loaded the deep learning function for the robot named Baxter, letting it sit on the high chair in front of the table like a baby in everyday objects piled up with a table.
Baxter is a modern two-arm industrial robot designed to perform repetitive tasks in a factory floor environment. Each of its arms consists of a standard two-finger parallel claw and an HD camera that lets the robot see what it is looking for. It also has a Microsoft Kinect sensor that provides an unobstructed view.
Pinto and Gunta are programming for Baxter, grabbing an item separately from other "neighbors", then randomly finding a point on the table, allowing the robot's two-finger parallel claws to rotate a certain angle and then directly grab Take this item. The robot then raises its arm and uses the force sensor to determine if the grip has been successful. This process is repeated 188 times and the angle of each transition is 10 degrees.
In order to make the robot learn better, Pinto and Gunta put a lot of items in front of Baxter's desk, and let it stay alone for 10 hours, without human intervention. If the robot drops the item on the floor, there are many alternative items on the table for it to practice without interruption.
Baxter's deep learning method is very standard, and the researchers have loaded it with a conventional neural network that has some basic object recognition skills before learning to grasp. However, there are still two network levels that require learning for random crawling experience.
The research team used a second level of learning to improve Baxter's skills. After picking up some basic items, they offered Buckster some new things, including what they had seen and new objects for it.
After more than 700 hours, Baxter tried 50,000 grabs for 150 items (including unsuccessful crawls). These items include TV remotes, many different plastic toys, and some similar items. This allows Baxter to predict whether the chances of success will be 80%.
Interesting research such as Baxter's interaction with the world will have an important impact. The key point is that Baxter is as easy to adapt as a messy, relatively unruly environment. More importantly, the skills to grasp items are basically done by self-study.
Of course, Baxter and its neural network are as flexible as a baby and have a long way to go. The next step in learning is to learn to grasp the strength so that you don't get rid of it when you get a vulnerable item.
Perhaps the ultimate test for Baxter would be to challenge the toothpaste – squeeze the bean-sized toothpaste onto the toothbrush. After all, this is a very important thing that human beings must learn from an early age.
Watch the video and learn to make cocktails
Industrial robots need to spend weeks reprogramming to perform a complex new task, which makes the reorganization of modern manufacturing lines very expensive and slow.
If the robot can see others do it before they are qualified for the new job, the whole process may be described as "suddenly advancing." This is the "point" in a project run by the University of Maryland, and they are educating robots to become "hardworking students."
Yang Yezhou, of the University of Maryland's Independence, Robotics and Cognition Laboratory, said: "We call it the 'Robot Training Academy'. We ask experts to demonstrate the task to the robot and let the robot do most of the work according to the task flow, then Fine-tuning to complete the task."
At a related conference in St. Louis, the researchers showed a robot that can modulate cocktails, and it was the above method that completed the modulation task. The guy with two industrial robots comes from a Boston-based company called "Rethinking Robots." A person pours liquid from several bottles and mixes them into a small drink in a small jar. The robot watches the whole process and copies all the movements. The key is that the order in which the bottles are picked up is correct.
This requires training a computer system to adapt to specific robotic actions. A recent paper by the research team pointed out that a robot learns to select different items and needs to use two different systems to watch thousands of teaching videos. One system is to learn how to identify different items, and the other system is clearly different. Type of gripping action.
Watching thousands of instructional videos may sound like a lot of time, but the learning process is much more efficient than re-programming the robot. What's important is that it makes it easy for the robot to "get started" with new tasks. The learning system for gripping movements includes an advanced artificial neural network, which has developed rapidly in recent years and is now being used to develop many types of robots.
Researchers are introducing their learning robots to several manufacturing companies, including e-commerce and automakers, to see if they can be applied to industrial production. These companies have also asked engineers to re-program robots to expand their robotic tasks, but most of them take a month and a half to get. Yang Yezhou said: "We can save at least half of our time with our method."
The project reflects two major trends in the robotics industry, one is to find new ways to learn robots, and the other is that robots can operate in the most human way.
Robots also use learning software
The rapid advancement of robotics is partly due to advances in hardware, including computer chips, sensors, and actuators, but software has also advanced technology. For example, the open source robotic operating system makes it easier for engineers to add new skills to the robot, so it doesn't have to start from scratch.
When kids are toddlers, they can quickly identify a delicious meal or avoid catching stinging. A software released recently allows robots to learn through experience rather than editing programs.
Advanced machine learning software, such as Brain OS, allows robots to see more advanced skills and also allows robots to learn more advanced technical experience.
Brain OS was developed by a "brain company" backed by Qualcomm, a mobile chip manufacturer based in San Diego, USA. Such software ultimately makes the robot easier to use. Users of the Brain OS can easily train the robot to do simple tasks, such as moving toward a specific object, without writing new code or accessing the graphical user interface.
For robots, working in a complex, constantly changing environment is very difficult, which is an important reason to turn learning ability into a reliable way. Some commercial robots have been able to do simple learning, but Brain OS software includes a set of tools for robot learning that make it easier for robotic engineers to access the software library for object recognition, navigation and operational tasks. program of.
At the robot conference held in Boston last year, the brain company showed the software, when the company used a red prototype robot to assemble a Segway electric balance car with two cameras for eyes, which showed an object to the robot. Then the robot can follow the object and never leave.
Brain OS's learning capabilities include providing information to virtual networks that mimic neurons and synapses, and then providing positive or negative feedback, a process known as "supervised learning." In recent years, this learning method has attracted much attention because it is very effective.
Todd Hilton, executive vice president of Brain, said in a statement that machine learning software is mostly aimed at academic researchers rather than industrial engineers. “Brain OS solves this problem by providing a centralized technical framework for commercial robotics applications that are close to prototypes.â€
The brain company also released a software version of the chip called bStem (short for brainstem) developed by Qualcomm. The chip can be used to design a way to simulate brain work, storing and processing data in parallel. Such "neuromorphic" chips can be used to efficiently run simulated neural networks, and Qualcomm is one of the companies that wants to commercialize this technology.
Concerned about surprises
Vehicle Hoses,Car Water Hose,Automotive Vacuum Hose,Exhaust Hose For Car
Changxing VACUFLEX Hose Technology Co., Ltd. , https://www.vacuflex-cn.com