Zhi DongXi (public number: zhidxcom)
Compile the | Zhihuai
Edit | Yunpeng
Zhidong April 11 news, has been the development of humanoid robots, is in the direction of reproducing human behavior, and then assisting human life. Recently, relevant researchers from the United Kingdom and Japan have released some new technologies, which can allow robots to have "the same sense of touch as humans" and enable robots to engage in more elaborate operations such as "helping humans wear clothes and peel bananas".
▲ "Artificial fingertips" are touching ball toys
Among these technologies, there are the results of repeated training of robots using "artificial neurons", and there are also new breakthroughs under the blessing of new elements such as "artificial fingertips". The real application of these technologies will bring new solutions to problems such as industrial manufacturing, medical care, and labor shortages.
First, 3D printing "artificial fingertips" gives robots a sense of touch comparable to real people
Contact is the most direct way for humans to perceive the world, and "touch" is an important part of the human nervous system, which plays an important role in the emotional interaction between humans. The skin and the nerves that cover it under the skin are the bridges that transmit "feelings" to the brain. Touch helps humans distinguish objects, and finally the brain controls the hands and chooses to hold their strength. But for robots made of metal and circuits, how can human "touch" be reproduced in them?
Researchers at the University of Bristol in the United Kingdom have come up with a solution – TacTip (artificial fingertip), also known as "light haptic sensor". Mandayam Srinivasan, who studies 'touch behavior' at the University of London, describes TacTip as: "Close contact between natural touch and artificial touch, which is a necessary step to improve the robot's 'touch scheme'." ”
▲ The robot equipped with TacTip successfully picked up a thousand paper cranes
The birth of the TacTip prototype dates back to 2009, when researchers at the University of Bristol hand-assembled the first "artificial fingertip" inspired by human skin. But subject to the technology of the time, the first "artificial fingertip" was not as slender as a real finger, it was the size of a canned soda. Then, with the rapid development of 3D printing technology, finally in 2018, the R & D team relied on 3D printing technology to compress the volume of "artificial fingertips" to the size of adult big toes.
At the same time, thanks to the ability of 3D printing technology to create a multi-layered structure similar to human skin for "artificial fingertips", the research and development team has recently integrated "artificial neural networks" into "artificial fingertips". At this point, the "complete body" of TacTip officially landed.
Engineers have long tried to make robots as dexterous as humans, equipped with "artificial neural networks" even though one of the solutions, but Mandayam Srinivasan, a touch researcher from the University of London, said: "At present, the feedback of robots' touch results is still much lower than that of humans." But TacTip's R&D team seems to have figured out a more ideal solution, which is the communication point between "human touch" and "machine response" - signals.
When the skin of the human fingertips comes into contact with the object, the nerve endings all over the skin transmit the command to the brain to "touch something" through the deformation of the synapse, and then the nerve will send two signals to the brain, "fast" signal to help us avoid the object falling, and the "lagging" signal is used to convey the shape of the object. I believe that everyone is about to see a rough idea, haptic feedback to the brain is in the way of neural signals, instruction feedback to the robot is in the form of digital signals, both of which are in the form of "signals", if you can accurately use digital signals to simulate neural signals transmitted to robots, then perhaps you can make robots have a sense of touch comparable to real people. The principle of TacTip is officially like this.
First of all, in order to make the signal simulation "accurate", the research and development team made a "rubber surface" like human skin for TacTip, and installed a set of needle-like raised arrays similar to human synapses under the "rubber surface", which simulated the middle epidermal ridge of human skin. The "antennae" that make up these arrays are tough and elastic, and when TacTip begins to "touch" the array that touches the surface of the object, it will begin to bend, generating a "fast signal" in human touch by the speed of bending, avoiding TacTip from holding the object unsteadily.
▲How TacTip works
Secondly, underneath this set of "arrays", TacTip's R&D team also installed a camera for it to monitor the degree of curvature of the "raised array", and the "degree of curvature" recorded by these cameras will be converted into a "slow signal" in the human sense of touch, allowing TacTip to judge what it "got".
This principle was endorsed by Sliman Bensmaia, a neuroscientist from the University of Chicago who studies the basics of touch neurons, who believes that most human touch is derived from skin mechanics, and the structure and method of TacTip fit this law.
When TacTip was officially completed, Nathan Lepora, an engineer from the University of Bristol, and his colleagues in the R&D team began to test TacTip for the first time, touching the object as a "corduroy material", because the surface texture of this material is staggered and the touch is changeable, and the final result will be judged by the reference sample of the neuronal signals of the real person's touch.
Happily, TacTip's test debut was outstanding, and based on the results published by its R&D team in the Journal of the Royal Society Interface on April 5, TacTip was able to distinguish the changes in the bulges of the "corduroy material" as accurately as the real person, as well as the gaps between the textures, and its output allowed neuronal signals to contact the "corduroy material" with the real person contact with the "corduroy material" The output is highly matched to the neuronal signal.
▲TacTip feedback touch neuron information schematic
But TacTip's shortcomings are also revealed - it is not as sensitive as the skin of a real person's fingertips. In simple terms, human fingertip skin can feel a gap similar to the width and narrowness of a pencil pellet, but to be "touched" by TacTip, this gap needs to be tripled. Research and development engineer Nathan Lepora believes that as long as the team can develop a thinner rubber surface layer that integrates more "raised arrays", the sensitivity of TacTip will be further improved as the density of "raised arrays" increases.
So before the second test began, the R&D team not only added more "bulging arrays" to the TacTip, but also added microphones to the previous structure. The microphone was added to collect the sound information generated by friction when the "bulging array" touches the surface of the object, in order to simulate another set of neural endings in the deep skin of humans that perceive "vibration", which gives TacTip the same ability to feel the roughness of the surface of the object as humans.
The "enhanced" TacTip thus ushered in the second test of the "enhanced version", in which the testers increased the difficulty and tried to make the TacTip distinguish between 13 different textile fabrics. Of course, the final result is still not disappointing, under the dual blessing of the camera and microphone, the neuronal signal output by TacTip is still comparable to the neuronal signal output when the real person touches.
TacTip's performance led Levent Beker, a mechanical engineer at The University of Cochi who works on wearable sensors, to exclaim that "the robot hand can finally feel the sensation of human fingers." Bensmaia, who had previously been a big believer in TacTip's principle, was even more praised after seeing the actual test results, and he felt that no one else had taken the most interesting way to give robots a "touch" like TacTip so far, which was very cool. At the same time, Bensmaia believes that the deformable TacTip can be directly attached to the robot's mechanical fingers or toes to help the robot detect, pick up and manipulate objects.
Commenting on this vision, Leporar of TacTip's R&D team said: "It is true that today's 'robot hands' have to rely on robotic arms and precise programming to get the job done, and they have a hard time grasping small and hard objects, such as toothbrushes and pens. TacTip, on the other hand, can allow robots or mechanical prostheses to handle objects of all shapes and sizes without relying on programming. ”
At the same time, for the future size of TacTip, Leporar also said that with the advancement of 3D printing technology, as well as the reduction of the size of camera and microphone components, TacTip will further become smaller and more refined, closer to the area of "human fingertips", while a smaller size allows TacTip to detect more fine textures.
In the face of Lepora's self-confidence, Bensmaia is conservative, he feels that in the final analysis, the tactile nerve signals simulated by TacTip are not exactly the same as those of the real fingertips, because the signal feedback of the real skin is stronger. "It's just infinitely closer to the real skin." Bensmaia thinks. And for the future, "how small" can TacTip become? He also said it was difficult to predict.
But in any case, the success of TacTip is gratifying, it is conceivable that it not only allows robots to have a "sense of touch", bringing them the ability to actively perceive and distinguish objects around them, but also makes robots free from the shackles of fixed programming instructions, giving them the ability to actively change the corresponding actions (such as grasping the force of things), making the robots themselves more "delicate".
Applied to the field of mechanical prosthetics, TacTip can reproduce the "touch of the fingertips", reopen the information transmission channels of the brain and nerve endings, and regain the ability of the disabled to touch things "completely". Robert Shepard, a materials scientist at Cornell University, said that fundamentally, TacTip's research helps scholars figure out how "touch" works in human nerves, and he believes that TacTip's creative team has basically recognized the principle of nerve endings feeding the brain back "touch", so it can allow the crane equipped with TacTip to pick up the paper plane "gently" without destroying it. For others, TacTip is a presence worth learning and understanding.
Second, help people dress, peel bananas on behalf of others, the smaller the action, the more delicate it is
Along with the "artificial fingertip", there are also related technological achievements of "robot successfully assisting medical mannequins to dress", and "Japanese robots can peel off complete bananas".
The results of "Robots Successfully Assist in Dressing Medical Mannequins" were jointly published on the science website by Fan Zhang and Yiannis Demiris. By experimenting on a medical mannequin, they successfully completed the automated process of "removing clothes from the hanger", "finding the patient on the bed", and finally "unfolding the clothes, lifting the arm of the mannequin, and finally dressing the model".
The success of this experiment means that patients who have lost the ability to move their upper limbs, such as high paraplegia, can successfully perform dressing operations with the help of robots, which will greatly save the labor cost of caring for patients for patients' families.
According to the R& D team, in order to achieve the process of allowing robots to assist humans in dressing, they have to face and solve two major challenges in the research and development process. First of all, it is necessary to let the robot change the clothes from the impenetrable state of "hanging" on the hanger to the dressable state of "taking" the clothes; secondly, it is to make the action of "dressing the patient" from "simulated instructions" to "mechanical action".
In response to the first problem, the R&D team chose a reasonable "pre-grasp" operation plan to solve. Before letting the robot "take clothes" action, they first let the robot "measure the distance between itself" and the hanger, and then let the robot know the distance between themselves and the hanger in advance, thus solving the problem of "fixed distance". For the second problem, the R&D team introduced the relevant algorithm of "clothing physics" for the scene simulator of the robot, and through the comparison with the neural network data, the observation of the real clothing was formed, and the physical similarity of the simulator was measured, so as to correct the error that the simulator might produce.
At present, the two-arm robot that adopts this technology has a dressing success rate of more than 90%. Imagine that when this technology continues to advance, and the two-arm robot can not only dress but also carry out operations such as feeding, covering quilts, and delivering things, then the help for the medical field will be unlimited.
Finally, let's look at the university of Tokyo researchers who successfully let a two-arm robot peel out a complete banana. Presumably everyone knows that the smaller the human amplitude of the action, the higher the difficulty of robot reproduction, and the higher the precision requirements for the mechanical structure, whether it is angle, strength, etc., it is a major test for the programming operation of the robot arm.
Reuters released information that the ISI laboratory of the University of Tokyo's School of Information has released a video, showing a two-arm robot that can completely peel off the peel of a banana in about three minutes without harming the pulp. Although the success rate is only 57%, it is undoubtedly of great significance for the study of "precision operation" of robots.
Members of the R&D team, Heecheol Kim, Yoshiyuki Ohmura and Yasuo Kuniyoshi, used a "deep simulation learning" approach to train the robot, demonstrating hundreds of banana peeling actions to generate enough data to allow the robot to repeat the learning, and finally after 13 hours of training, the robot successfully completed the "banana peeling" action.
▲ The two-arm robot picks up the banana and peels it without squashing the fruit inside
While training the robots to learn more abilities through this method, Kuniyoshi on the team believes that their training method can effectively allow the robots to learn more elaborate human movements, such as repetitive assembly line work, and solve the problem of Labor Shortage in Japan.
It can be said that whether it is dressing or peeling bananas, the work that robots can do today has become more refined and complex, and the corresponding application scenarios of robots in our lives are also increasing. From taking care of patient radiation to solving the problems of the entire medical industry, from replacing workers radiation to solving the productivity problems of the whole society, the success of these technologies is likely to drive the process of society in the future. Successfully allowing robots to touch, dress and peel bananas is also an external manifestation of technology accumulation.
Conclusion: Reinforcement learning has become the key, and "human-like" robot technology has ushered in a new breakthrough
From the above, it is not difficult to see that repeated intensive training is still the main way for robots to learn, but the addition of human biomimetic technology such as "artificial fingertips" provides a new way for robots to learn. Obviously, "artificial fingertip" is a unique existence in the current robot industry, but relatively speaking, compared with the traditional "simulation training", the mechanical structure of "artificial fingertip" is more precise, and the technology involved in the product itself is more and more complex.
Robots should not only be the carrier of instructions, but also have the intelligence of the human brain, have the ability to learn, understand problem analysis and can perform precise operations. The success of the three technologies of "artificial fingertips", assisted dressing and banana peeling corresponds to the above three conditions, and also reflects the advancement of artificial intelligence technology from the side. For the industry, in the future, the training method of robots is likely to change from the original "mechanical imitation" to "active learning" through the blessing of sensors, and once such a change occurs, the process of robot learning humans will be greatly accelerated.
Source: Science, Reuters