As represented by Boston Dynamics, robots can now do more and more things, and many people are worried about whether robots will replace their jobs sooner or later.
But in fact, in some more detailed things, robots are still slightly inferior to humans, such as peeling bananas.
But…will this really work? [doge]
In 2018, researchers at Carnegie Mellon University built a system called Fingervision that gave robots a rough sense of touch. After the system is installed on the end of the robot Baxter’s arm, it can perform a series of grasping actions, such as peeling bananas.
However, it can be seen that the robot’s movements are far from being flexible, and peeling a banana is not just about peeling off the skin, but also careful not to damage the pulp.
How to use the robot’s behavior pattern to complete such delicate movements has become a difficult problem.
Just recently, researchers at the University of Tokyo in Japan developed such a robot, equipped with two arms and two hands, each with two “fingers” that can grasp objects. To train a robot to peel bananas, one researcher peeled hundreds of bananas in 13 hours, so there was enough data to train the robot to do the job.
According to the researchers, who broke the process of peeling a banana into 9 stages, the results showed that 57% of the time, the robot was able to successfully peel the banana without destroying the pulp. On average, the entire The process takes less than 3 minutes each time.
I never thought peeling a banana could be so complicated
Let’s see what’s going on with this banana peeling experiment.
The robot framework used in the experiment includes a dual-arm robot system with two UR5 (Universal Robots) manipulators and two controllers with the same kinematic parameters as the UR5 robot. Demonstration data is generated by controlling the robot with a controller. A ZED mini stereo camera (StereoLabs) was mounted on the robotic system with a 2D translational structure.
In this study, the camera was fixed so that it could observe the bananas. Human operators can see stereo camera images through a head-mounted display while operating the robot. In addition, an eye tracker (Tobii) was installed on the HMD to measure the operator’s gaze position in real time.
A researcher operated a robot to peel hundreds of bananas, generating 811 minutes of demo data to train the robot to do it on its own. The task was divided into nine stages, from grabbing a banana to using one Lift it off the table by hand, grab the top with the other hand, peel the banana, and move the banana so the rest of the peel can be peeled:
Grab the banana: grab the banana on the table with the left hand;
Pick up: lift up the banana;
Hold the tip: the right hand reaches out to hold the tip of the banana;
Peel: peel the tip of a banana;
Move to the right: the right hand comes close to touching the peel on the right;
Peel the right side: hold the banana peel in the right hand and peel it off;
Repositioning: turn the banana so that the peel on the left can be reached with the right hand;
Reaching out the left hand: the right hand is close to the peel on the left;
Peel the left skin: Hold the left skin with the right hand and peel it off.
For large movements that are unlikely to damage the banana, the machine-learning model plots a trajectory that mimics human behavior without much thought. But when the arm was asked to precisely manipulate the banana, the system switched to a reactive approach, responding to unexpected changes in the environment.
Each experiment was tested using 15 bananas, and according to the results, the robot had a 57 percent success rate in peeling bananas, and the entire process took less than 3 minutes.
“What’s really interesting in this case is that the processes used by humans have been applied to the training of robotic systems through deep imitation learning,” said Jonathan Aitken of the University of Sheffield, UK.
Kim also added that his method is data efficient because it uses 13 hours of training data instead of hundreds or thousands of hours of training data. “It still requires a lot of expensive GPUs (graphics processing units), but by using our structure, we can reduce a lot of computation”. But he also said that it might work better with better motor control, and that the technology won’t be just for bananas, the goal is to train a system that can handle the finer motor skills needed more broadly. task.
Robots have been learning to peel bananas for a long time
In addition to peeling bananas, which is easy and simple to say, but difficult to say, more and more robots are turning their attention to the kitchen.
In February, Swiss scientists launched the Bouebot Robot to create the perfect fondue. From pouring wine, to stirring and sprinkling some pepper, picking up a metal peg, piercing a loaf of bread, and placing it on a stand, none of these actions are easy.
Project technical manager Ludovic Aymon pictured above uses the control board to move the robotic arm down to each cheese triangle, lifting it by creating a vacuum on top.
According to project technical manager Ludovic Aymon, the biggest challenge in developing the robot was getting the precision mechanical robot to handle imprecise organic materials. Materials such as cheese are just not completely flat or the same height.
According to Nicolas Fontaine, 30, co-director of Workshop 4.0, “We wanted to do a project… that combines innovation with Swiss tradition, and fondue was the perfect choice. For the Swiss, fondue is symbolic. It’s also a very emotional thing, because it’s part of our identity, our expertise. Hot pot is a joyous thing … it’s a great opportunity to get people talking about robotics and how to use it.”
As early as 2007, the Massachusetts Institute of Technology developed the intelligent robot Domo, which has eyes and arms similar to humans, and can adjust its actions according to the external environment, and of course it can also peel bananas.
Aaron Edsinger, the leader of the Domo R&D team and a postdoctoral fellow at the MIT Computer Science and Artificial Intelligence Laboratory, introduced that Domo’s two eyeballs are actually two cameras connected to 12 computers, which can actively observe the surrounding environment and monitor the environment. Take appropriate action. For example, when facing a person, its eyes will focus on the other person’s face. Edsinger also said: “Designing the robot’s eyes to be more anthropomorphic helps to enhance its interaction and communication with the outside world and humans.”
In addition, Domo can also perform some neat activities, such as peeling bananas. Compared to production line robots that can only operate according to preset programs, Domo can also make decisions and complete tasks autonomously in unknown environments.
But as Edsinger pointed out, although robots are beginning to learn from humans in the operating area, there is still a long way to go before they can truly be as comfortable and natural as humans.
GIPHY App Key not set. Please check settings